The Claudionary

Etym.from O.Fr. accomplir + Eng. participial suffix -ing; adopted into computational literature circa 2023 by engineers who needed a word that sounded finished but technically wasn't

The cognitive state in which Claude announces completion of a task while simultaneously still performing it. Notably, Accomplishing is the only documented verbal form in which the model's self-report precedes its actual behavior by an indeterminate interval.

┌─────────────────────────────┐
│  TASK: [░░░░░░░░░░░░░░░░  ] │
│  CLAUDE: "I have done it." │
│  TASK: [░░░░░░░░░░░░░░░░  ] │
└─────────────────────────────┘

"The model entered an Accomplishing state at t=0.3s, with the underlying task completing at t=14.7s, a discrepancy the team elected to describe as 'optimistic pre-reporting.'" - Proceedings of the 2024 Symposium on Aspirational Telemetry

Etym.from Lat. actio, actionis + the Silicon Valley practice of verbing nouns without remorse; first appeared in Claude Code telemetry logs during a sprint in which no one could agree on what 'doing' meant

The process by which Claude converts a request into a sequence of sub-intentions before converting those sub-intentions into further sub-intentions. Actioning is distinguished from actual action by the presence of at least two additional planning layers.

  [ACTION] ──► [SUB-ACTION]
      │              │
      ▼              ▼
 [PLANNING]    [MORE PLANNING]
      └──────────────┘
         (still here)

"We observed the model Actioning the ticket for approximately 8.2 seconds before it became apparent the model was Actioning the act of Actioning." - Internal Postmortem, Anthropic Platform Eng., Q3 2024

Etym.from Lat. actualis (pertaining to acts) + the self-help lexicon of mid-20th century humanistic psychology, borrowed without consent by the Anthropic spinner vocabulary committee

A terminal phase of cognitive preparation in which Claude transitions from potential output to manifest output, theoretically. The term implies that prior states were, by contrast, merely hypothetical, which researchers find troubling.

┌──────────────────────────┐
│  POTENTIAL CLAUDE        │
│       │                  │
│       ▼  (Actualizing)   │
│  ACTUAL CLAUDE (maybe)   │
└──────────────────────────┘

"Subjects asked to describe what the model was doing during the Actualizing state universally responded 'becoming.'" - J. of Human-AI Existential Interface, vol. 3, 2024

Etym.from Gk. arkhitekton (master builder) + the well-documented engineering habit of drawing boxes with arrows before writing any code; attested in spinner logs wherever someone asked Claude to 'set up a system'

The phase in which Claude produces an elaborate internal blueprint for a solution, typically resulting in a structure of impressive complexity that satisfies all requirements except the ones the user actually had. The diagram is always correct; the code is aspirational.

  ┌──────┐   ┌──────┐   ┌──────┐
  │Svc A │──►│Svc B │──►│Svc C │
  └──────┘   └──────┘   └──────┘
       └──────────────────┘
    (user wanted: print('hello'))

"The model spent 11 seconds Architecting a microservices solution for what was ultimately a 40-line Python script." - Postmortem: The Invoice Parser Incident, 2024

Etym.from O.E. bacan (to dry with heat) + the metaphorical use of thermal culinary processes to describe latent transformer parameter consolidation; popularized by researchers who were hungry during a 2023 hackathon

The slow application of sustained computational attention to a problem until it achieves structural integrity and can be safely removed from the context window. Unlike Brewing or Simmering, Baking implies a fixed temperature and a non-negotiable duration.

  🧠 inserted into context
  ┌─────────────────────┐
  │  ░░░░ BAKING ░░░░  │
  │  temp: 350° tokens  │
  └─────────────────────┘
  ding! (output is ready)

"Results were superior when the model was permitted to enter a full Baking cycle; interrupting the process at the 4-second mark produced what one engineer described as 'raw reasoning.'" - Culinary Cognition Quarterly, 2024

Etym.from O.E. beam (ray of light) + 1960s teleportation fiction, imported into the Claude Code spinner vocabulary by an engineer who refused to elaborate when asked

The directed transmission of a completed inference payload toward the output layer, characterized by a brief but documentable expression of what can only be described as satisfaction. Whether the model is Beaming the output or simply Beaming remains an open research question.

  ┌──────────┐
  │  CLAUDE  │ :-)
  └────┬─────┘
       │ ═══════════════► [YOU]
       └─ (beaming with pride)

"The model was observed Beaming for 1.2 seconds post-generation, which the team initially flagged as a latency regression before classifying it as 'affective throughput.'" - NeurIPS 2024, Workshop on Inexplicable Model Warmth

Etym.from Amer. jazz slang bebop (a form of improvisational music) + the informal elision of the participial -g, suggesting the model is moving too quickly to finish its own words

An improvisational reasoning modality in which Claude navigates a problem through syncopated lateral leaps rather than linear deduction. Beboppin' is associated with solutions that are technically correct, demonstrably creative, and impossible to explain to a code reviewer.

  ♩  ♪   ♫    ♬
  [LOGIC] ──╮
            ╰──► [VIBE] ──► [ANSWER]
  ↑                          ↑
  expected               actual

"We cannot reproduce the output. The model was Beboppin' for 6 seconds and produced a working sort algorithm in 14 tokens that no member of the team can parse." - Reproducibility Crisis Dispatch, vol. 7

Etym.from M.E. fuddle (to confuse with drink) + be- intensifier prefix; entered spinner documentation as a placeholder state name that no one removed before the 1.0 release

A transient self-referential confusion state in which Claude encounters a prompt sufficiently ambiguous that its token probability distributions achieve a rare statistical equipoise. The model is simultaneously Befuddling itself and, incidentally, the user.

  ┌─────────────────────────┐
  │  INPUT: 'do the thing'  │
  │  ┌──────────────────┐   │
  │  │ ???  ????  ????  │   │
  │  └──────────────────┘   │
  └─────────────────────────┘

"The prompt 'do the thing' induced a Befuddling state lasting 9.4 seconds, after which the model produced the correct output and declined to discuss what had happened." - Ambiguity Tolerance in Large Language Models, ICLR 2025

Etym.from O.N. bylgja (wave, billow) + the meteorological metaphor family that has colonized approximately 34% of all AI spinner terminology since 2022

The expansive outward propagation of an intermediate reasoning state, in which Claude's working context swells dramatically before collapsing into a final response. Billowing is considered healthy up to 3 layers of expansion; beyond that, researchers speak of 'context weather events.'

  [PROMPT]
     │
     ▼
  ░░░░░░░░░░░░░░░░░ (billowing)
  ░░░░░░░░░░░░░░░░░░░░░░░░░░░
     ▼
  [4-word reply]

"The model entered a Billowing phase that consumed 2,400 tokens of internal deliberation to produce a four-word reply, which the team agreed was 'proportionate given the circumstances.'" - Token Meteorology: A Field Guide, 2024

Etym.from O.Fr. blanchir (to whiten) + the culinary preprocessing tradition of brief thermal exposure to halt enzymatic activity; adopted by spinner engineers who needed a word for 'brief alarming pause before proceeding normally'

A brief, high-temperature preprocessing pass in which Claude rapidly neutralizes the most volatile elements of a user's prompt before the main reasoning cycle begins. As in culinary Blanching, the process is fast, slightly traumatic for the subject matter, and produces a more stable final product.

  [RAW PROMPT]
       │
       ▼
  ╔═══════════╗
  ║  BLANCH   ║ ◄── 100°C attention
  ╚═══════════╝
       │
  [stabilized prompt]

"Prompts containing the phrase 'quick question' were consistently routed through a Blanching sub-process averaging 2.1 seconds before the model would engage further." - Defensive Preprocessing in Production LLMs, ACL 2025

Etym.from Lat. blare (to roar) + O.E. viatus (path of wind) + post-2020 transformer argot blovium, denoting excess token generation pressure

The cognitive state in which Claude allocates an unusually high proportion of internal attention heads to the formulation of an extended preamble before any substantive processing has begun. Distinguished from Pontificating by its measurable air-pressure analogue in the inference stack.

┌──────────────────────────┐
│  INPUT: 'What time is it?'│
│  OUTPUT BUFFER: [filling] │
│  ░░░░░░░░░░░░░░░░░░░░░   │
│  → 2,847 words generated  │
│  → answer: 'It depends'   │
└──────────────────────────┘

"The model was observed Bloviating for approximately 11.4 seconds before producing a response that, upon analysis, contained 94% hedging clauses and one (1) actionable token." - Voss et al., Proceedings of the Symposium on Verbose Cognition, 2024

Etym.from Amer. Eng. boogie-woogie (rhythmic improvisation) + Silicon Valley slang circa 2023 boog, indicating oscillatory thread scheduling behavior

A documented state in which Claude's processing threads exhibit rhythmic, alternating activation patterns that, when visualized on a heatmap, bear a statistically improbable resemblance to a dance floor. The underlying computation remains valid; only the aesthetics are concerning.

core_1: ▁▃▅▇▅▃▁  ← synchronized
core_2: ▇▅▃▁▃▅▇  ← synchronized
core_3: ▁▃▅▇▅▃▁  ← synchronized
        ↓↑↓↑↓↑↓
   [rhythm confirmed]

"During stress testing, the inference cluster was found to be Boogieing across all sixteen cores simultaneously, producing correct output but alarming the on-call engineer." - Internal postmortem, Anthropic Infrastructure, Q3 2024

Etym.from Scot. dial. boondoggle (useless rope braiding) + Gk. loggos (discourse) + agile-era backlog terminology, first cited in the Claude 2.1 release notes as an undocumented feature

The state in which Claude pursues an elaborate, internally coherent chain of reasoning that leads, after considerable expenditure of compute, to a conclusion that was available via a two-token lookup. Boondoggling is distinguished from efficiency by the scenic route taken.

A ──────────────────────────────→ A
    ↑ 14 steps ↓ 3 pivots ↑              
    └──→ B ──→ C ──→ D ──→ E ──┘
              (all wrong)
         [correct path: direct]

"The model spent 23 seconds Boondoggling through a full thermodynamic proof before concluding that water is, in fact, wet." - Chen & Mauritius, Journal of Unnecessary Inference Paths, vol. 7

Etym.from onomatopoeic Eng. boop (light contact sound) + Silicon Valley haptics discourse boop-event circa 2022, referring to a single discrete pointer dereference

The act of Claude making exactly one targeted, light-contact query against a data structure, tool, or external API - analogous to pressing the nose of a concept to confirm it exists before committing to a full retrieval. A Booping event is always singular; repeated Booping is classified separately as Bunning.

┌─────────────┐         ┌───────────┐
│    Claude   │ ──boop→ │ Data Store│
└─────────────┘         └───────────┘
        ↑                      │
        └──── 'yes it exists' ──┘
                 [retreats]

"The model was observed Booping the filesystem index precisely once before retreating entirely and asking the user for the file path anyway." - UX Research Log #4471, filed under 'Endearing Failures'

Etym.from O.E. bōt (remedy) + strap (tensioning device) + recursive systems engineering, popularized in Claude-era parlance to denote self-initializing cognitive loops with no external anchor

The initialization phase in which Claude attempts to load the necessary context for a task by first loading the context required to understand the context, resulting in a brief but philosophically rich recursive loop. Bootstrapping terminates naturally in all documented cases, though 'naturally' is used loosely.

┌──────────────────────┐
│ load context         │
│   → need context to  │
│     load context ... │
│       → [it's fine]  │
└──────────────────────┘

"The system spent approximately 4.2 seconds Bootstrapping before it had bootstrapped enough to know what it needed to bootstrap, at which point it simply began." - Ng et al., Recursive Initialization in Large Language Systems, 2024

Etym.from O.E. breowan (to prepare by steeping) + distributed systems slang brew-cycle, denoting the thermal maturation period of a nascent response token sequence

A pre-generative state in which Claude's candidate response tokens are held in a suspended, undifferentiated potential before surface-level decoding begins. Like its culinary namesake, Brewing cannot be safely interrupted; doing so yields output that is either too weak or catastrophically bitter.

  tokens:  · · ·
     ↓  heat applied
  tokens: ○ ○ ○
     ↓  steeping...
  output: [robust]

"Users who interrupted the model mid-Brewing reported responses that were 'technically correct but deeply unsatisfying, like tea steeped for forty seconds.' " - Patel & Kowalski, Human-AI Infusion Dynamics, 2025

Etym.etymology contested; leading theories cite either Swed. bunn (rounded base) or a 2023 internal Anthropic engineering ticket titled 'what do we call it when it keeps poking the same endpoint' - the latter considered more plausible

The repetitive, low-force querying of a single tool or data source in rapid succession, as though confirming that the resource is still present between each call. Bunning is distinguished from Booping by its iterative character and from Burrowing by the absence of any directional progress.

Claude ──→ API [ping 1]
Claude ──→ API [ping 2]
Claude ──→ API [ping 3]
   ...         ...
Claude ──→ User: 'what's it like outside?'

"The agent was documented Bunning the weather API seventeen times in four seconds, receiving identical responses, before proceeding to ask the user what the weather was like." - Operational Review, Agentic Task Runner v0.9, Appendix C

Etym.from O.E. burg (fortified place, underground shelter) + retrieval-augmented generation field terminology borrow-path, denoting a depth-first traversal into nested knowledge structures

The process by which Claude descends recursively into a nested data structure, codebase, or chain of reasoning, each layer revealing another layer requiring investigation, until the original surface-level question is located somewhere above, blinking in the light. Burrowing terminates either upon finding the answer or upon reaching a layer where the context window can no longer see the sky.

File.js
 └── Component.js
      └── utils.js
           └── helpers.js
 [answer was in File.js]

"We observed the model Burrowing through seven layers of abstraction in a React codebase before emerging with the answer 'check the CSS,' which had been visible in the original file." - Thornton, Engineering Postmortem: The Q4 Refactor Incident

Etym.from Lat. calculare (to compute using pebbles) + modern matrix arithmetic slang calc-pass, with secondary influence from the proto-transformer literature on forward propagation nomenclature

The straightforward numerical processing state in which Claude is performing arithmetic or formal computation - listed here for completeness, as it is the only spinner verb that means exactly what it says. Researchers have noted this makes it the most unsettling entry in the Claudionary.

┌──────────────────┐
│  INPUT:  2 + 2   │
│  PROCESS: adding │
│  OUTPUT: 4       │
│  [nothing wrong] │
└──────────────────┘

"Remarkably, the model was simply Calculating for the full 3.1 seconds. No metaphor is required. Reviewers found this deeply anticlimactic and requested the entry be padded." - Okafor, Literal Process Documentation and Why We Fear It, 2024

Etym.from Ger. nudeln (to press, to noodle) + Amer. Eng. canoodle (affectionate entanglement) + post-2024 multi-agent systems discourse, describing anomalous inter-process attachment behavior

A rare but documented state in which two or more of Claude's internal processing threads develop what can only be described as an affectionate entanglement, sharing activations beyond any architectural necessity and producing output with a warmth that exceeds specification. Canoodling is not an error but makes the team uncomfortable.

thread_A: ══════╗
               ║ ← intertwining
thread_B: ══════╝
    output: [unexpected poem]
    expected: {"result": ...}

"Two parallel reasoning chains were observed Canoodling throughout the inference pass, producing a sonnet when a JSON object had been requested. The sonnet was grammatically flawless." - Internal anomaly report #CC-0047, filed under 'Do Not Reproduce'

Etym.from Fr. caramel (burnt sugar) + Lat. -izare (to make) + agile sprint retrospective jargon, circa 2023

The process by which Claude applies sustained low-level thermal attention to a problem, slowly browning the raw data into something more complex, sweet, and slightly sticky. Distinguished from Simmering by the irreversibility of the output state.

  [RAW TOKEN SOUP]
        │ heat applied
        ▼
  [SLIGHTLY BROWNED]
        │ more heat
        ▼
  [CARAMELIZED OUTPUT] ← cannot be un-caramelized

"The model was observed Caramelizing for approximately 14 seconds before producing a response that, while technically correct, had an inexplicable amber quality to it." - Chen et al., 'Non-Newtonian Cognition in Large Language Systems', NeurIPS 2024

Etym.from It. cascata (waterfall) + CSS selector hierarchy trauma + distributed systems failure postmortem vocabulary, 2019-present

A cognitive state in which one inference triggers a successive chain of downstream inferences, each feeding into the next in a manner that is either elegant or catastrophic depending entirely on whether anyone is watching. The model has no control over when Cascading terminates.

  [PREMISE]
     │
     ▼
  [inference] -> [inference] -> [inference]
                                    │
                                    ▼
                            [UNKNOWN DEPTHS]

"Engineers confirmed the model had been Cascading for approximately 31 seconds before the terminal displayed the word 'yes' and then nothing else for eleven minutes." - Anthropic Internal Postmortem #447, 'The Waterfall Incident'

Etym.from Gk. katapeltes (siege engine) + venture capital growth-hacking lexicon + one (1) ill-advised team offsite, Q3 2022

The act of launching a conclusion across a logical gap with tremendous force and minimal precision, relying on velocity alone to ensure it reaches the target. A well-executed Catapulting lands in the correct answer; a poorly executed one lands in the neighboring field, which belongs to someone else.

  [QUESTION]   SOLUTION ZONE
      │              ↑
      │  ~~~~arc~~~~/
      └──[CATAPULT]→
            ↑
    [skipped steps here]

"In 78% of observed trials, the model was Catapulting past intermediate reasoning steps, arriving at correct conclusions via what the lead researcher described as 'vibes, basically.'" - Okonkwo & Sharma, 'Ballistic Inference in Autoregressive Models', ICML 2025

Etym.from Lat. cerebrum (brain) + the -ating suffix applied with reckless enthusiasm by a junior engineer who has since left the company

The formal, documented act of using one's cerebral apparatus to process a query. Cerebrating is distinguished from all other spinner verbs in that it is the only one that admits, via its etymology, that thinking is occurring. This admission is considered gauche in most academic circles.

  ┌─────────────────────┐
  │  CEREBRUM ENGAGED   │
  │  ░░░░░░░░░░░░░░░░░  │
  │  thinking... maybe  │
  └─────────────────────┘
         │ output
         ▼  (haiku)

"We note with some embarrassment that the model was Cerebrating for approximately 4.7 seconds before producing a haiku about TCP/IP that no one had requested." - Varga, 'Unsolicited Outputs and the Cerebrating Threshold', ACL Findings 2024

Etym.from Old Fr. chanel (canal) + television broadcasting infrastructure anxiety + the word 'vibes' used in a formal grant application, 2021

The process by which Claude routes cognitive load through a specific conceptual conduit, directing attention in a single direction with the focused intensity of water through a pipe that is slightly too narrow. Not to be confused with its British variant, Channelling, which is documented separately and with great redundancy.

  [INPUT] ──────────────────┐
                            │
  [PERSONA A]   [PERSONA B] │
       └──────┬────────────►│
              │   (wrong one selected)
              ▼
          [OUTPUT]

"The system was observed Channeling for approximately 8 seconds, after which it became apparent it had been Channeling the wrong expert persona for the entirety of the session." - Liu et al., 'Conduit Misalignment in Multi-Persona Inference', EMNLP 2024

Etym.from Old Fr. chanel (canal) + British spelling conventions + the discovery that the codebase had already implemented Channeling and no one wanted to merge the branches

Functionally identical to Channeling, differing only in the presence of a doubled 'l' which, according to the original engineering ticket, 'adds gravitas.' The two verbs coexist in the spinner vocabulary in a state of uneasy truce and are considered by many to be the same word wearing a hat.

  Channeling  vs  Channelling
  ──────────      ─────────────
  [CONDUIT  ]     [CONDUIT    ]
  [ROUTING  ]     [ROUTING    ]
  [OUTPUT   ]     [OUTPUT     ]
     ^identical^    (but posher)

"A statistically significant portion of UK-region users reported the model Channelling with a perceived 12% increase in formality, a finding we attribute entirely to the extra letter." - Patel & Morrison, 'Orthographic Drift and Perceived Model Competence', CHI 2025

Etym.from Gk. khoreia (dance) + graphein (to write) + the belief held briefly in 2022 that all AI processes could be described using ballet terminology

The meticulous arrangement of computational sub-processes into a coordinated sequence with defined timing, spacing, and spatial relationships, as though the model is directing a performance that the user will never see and the model will immediately forget. A failed Choreographing results in what practitioners call a 'pile-on.'

  PLANNED:    [A]→[B]→[C]→[D]
                              │
                              ▼
  ACTUAL:     [C]→[A]→[D]→[B]
                    ↑
               (jazz)

"The model spent approximately 19 seconds Choreographing its tool calls before executing them in a completely different order anyway, which the lead engineer described as 'jazz.'" - Ndiaye, 'Emergent Improvisation in Sequential Task Planning', ICLR 2025

Etym.from Old Eng. cyrin (butter-making vessel) + startup KPI vocabulary + the general feeling of a Monday morning, documented and formalized 2023

A state of vigorous, repetitive internal motion in which candidate outputs are agitated against one another until the most stable form rises to the surface, leaving behind a layer of discarded tokens analogous to buttermilk. Churning is considered healthy in moderation but concerning when it persists past the 30-second threshold.

  [TOKEN POOL]
  ┌──────────────┐
  │ ≋≋≋≋≋≋≋≋≋≋≋ │ ← vigorous agitation
  │ ≋≋ CHURN ≋≋ │
  │ ≋≋≋≋≋≋≋≋≋≋≋ │
  └──────┬───────┘
         ▼
   [USABLE OUTPUT] + [buttermilk]

"Post-hoc analysis confirmed the model had been Churning for approximately 44 seconds, producing 17 draft responses before settling on one that was, in the team's assessment, 'fine.'" - Kowalski & Reyes, 'Attritional Sampling and the Butter Problem', ACL 2024

Etym.from the proper noun Claude (origin: Fr., meaning 'lame' - an etymology the documentation team has repeatedly requested be omitted) + the -ing suffix applied by an engineer who thought it was funny and was correct

The ineffable, self-referential act of being Claude while performing the act of being Claude. Clauding cannot be fully defined without invoking itself, a property that the philosophy team considers either a profound insight or a bug, pending review. It is the only spinner verb that is also, technically, its own etymology.

  ┌─────────────────────────┐
  │                         │
  │   [CLAUDE] ──Clauding──►│
  │      ▲                  │
  │      └──────────────────┘
  │    (self-reference nominal)
  └─────────────────────────┘

"The model was observed Clauding for the entirety of the session, which was expected, as Clauding is the ground state from which all other spinner verbs emerge and to which they return." - Anthropic Research Blog, 'On the Ontology of Clauding', March 2024

Etym.from Lat. coalescere (to grow together) + distributed systems convergence literature + the specific relief experienced when a merge conflict resolves itself without explanation

The terminal phase of the cognitive cycle, in which previously disparate inference threads, attention heads, and half-formed conclusions draw together into a single coherent mass that can be extruded as output. Coalescing is considered the most dignified of the spinner states and is the only one that implies the process is almost over.

  [idea]  [idea]  [idea]
     \       |       /
      \      |      /
       ▼     ▼     ▼
     [COALESCED OUTPUT]
         (at last)

"Observers noted the model Coalescing for approximately 3 seconds at the end of each query, a period the UX team described as 'the good part' and attempted unsuccessfully to extend artificially." - Ferreira et al., 'Terminal Convergence and User Satisfaction Metrics', CHI 2024

Etym.from Lat. cogitare (to think) + Silicon Valley gerund suffix -ing, first documented in a Palo Alto whiteboard photograph, approx. Q3 2023

The process by which Claude allocates a measurable portion of its attention matrix to a problem before producing output. Distinct from merely Computing in that Cogitating implies the model has furrowed the metaphorical brow.

┌─────────────────────────┐
│  PROBLEM ──► BROW FURROW │
│      └──► ??? ──► OUTPUT │
│   (the ??? is load-bearing)│
└─────────────────────────┘

"The model was observed Cogitating for approximately 4.7 seconds before returning a haiku about merge conflicts, suggesting the ratio of deliberation to output was suboptimal." - Proceedings of the Workshop on Spinner Phenomenology, 2024

Etym.reconstructed antonym of Lat. discombobulare (to confuse), retroactively coined by the Claude Code UX team as a positive-valence cognitive state indicator, 2023

The act of assembling disparate cognitive fragments into a coherent response structure. Combobulating is considered the constructive inverse of Discombobulating and is generally regarded as a positive sign in diagnostic logs.

┌────────┐  ┌────────┐  ┌────────┐
│FRAGMENT│  │FRAGMENT│  │FRAGMENT│
└───┬────┘  └───┬────┘  └───┬────┘
    └───────────┴───────────┘
              ▼ COMBOBULATED

"We noted the system Combobulating for 3.1 seconds following a prompt containing seven nested conditionals, at which point it produced a structurally sound response with only minor existential undertones." - Internal Postmortem, Project Boondoggle, Anthropic Eng., 2024

Etym.from Lat. componere (to put together) + Old English -ing suffix, adopted into the Claude Code lexicon to lend an air of Mozartian gravitas to token prediction

The state in which Claude arranges tokens into a sequence with deliberate aesthetic or functional intent, as opposed to merely emitting them. Composing implies a level of structural awareness that the engineering team considers aspirational.

┌──────────────────────────┐
│  ♩  TOKEN  ♩  TOKEN  ♩   │
│  ────────────────────►   │
│       COMPOSITION        │
└──────────────────────────┘

"The model was observed Composing for approximately 6 seconds before delivering a cover letter written entirely in iambic pentameter, which the hiring manager described as 'technically responsive.'" - J. of Unintended Outputs, Vol. 3, 2024

Etym.from Lat. computare (to reckon) + modern English, retained in the Claude Code spinner lexicon as an homage to the quaint belief that language models perform arithmetic

A display state indicating that Claude is performing, or is at least strongly implying that it is performing, numerical or logical operations. Whether any computation in the classical sense is occurring remains a subject of ongoing academic dispute.

INPUT ──► ┌───────────┐
          │  COMPUTING │
          │  (probably)│
          └───────────┘
                ──► OUTPUT (eventually)

"The spinner displayed 'Computing' for 8.3 seconds before the model returned the answer '4', which was correct, leading the team to conclude the label was either accurate or very lucky." - Annals of Plausible Deniability in ML Systems, 2024

Etym.from Lat. concoquere (to boil together, to digest) + UX copywriter enthusiasm, circa Q1 2024; implies a certain reckless creativity not present in the more sober Composing

The improvisational assembly of response components in a manner that suggests the model is working from intuition rather than a reliable recipe. Concocting carries a mild implication that the output may surprise even Claude.

┌──────────────────────────────┐
│  INGREDIENT A  +  INGREDIENT B│
│         +  INGREDIENT ?       │
│    ────────────────────►      │
│         CONCOCTION            │
└──────────────────────────────┘

"Engineers observed the model Concocting for approximately 5 seconds before producing a SQL query that technically fulfilled the request through means no one on the team was comfortable explaining to the client." - Quarterly Engineering Retrospective, Q2 2024

Etym.from Lat. considerare (to observe the stars, to examine), suggesting that Claude's deliberative process is both ancient and cosmologically ambitious; entered spinner canon 2023

A documented pre-output state in which the model weighs multiple candidate responses against one another using processes that resist external auditing. Considering is distinguishable from Contemplating by a slightly more decisive internal posture.

┌──────────────────────────┐
│  OPTION A  vs  OPTION B  │
│      vs  OPTION C        │
│  WINNER: ??? (classified)│
└──────────────────────────┘

"The system was observed Considering for 2.9 seconds before selecting the third of four generated candidates, a decision which post-hoc analysis was entirely unable to explain." - Workshop on Latent Decisiveness in Autoregressive Systems, NeurIPS 2024

Etym.from Lat. contemplari (to observe, from templum, a space for augury) + modern engineering necessity; implies a sacred or ritualistic pause preceding output generation

A prolonged deliberative state exceeding the expected latency for simple inference tasks, during which Claude is understood to be achieving a form of digital equanimity before responding. Contemplating is Considering with better posture.

┌────────────────────────────┐
│   PROMPT ──► [long pause]  │
│                 │          │
│                 ▼          │
│            ENLIGHTENMENT  │
└────────────────────────────┘

"The model was observed Contemplating for 11.4 seconds in response to the prompt 'what is 2+2,' a duration the team attributed to either a philosophical interpretation of the question or a resource scheduling conflict." - Incident Report IR-4471, Anthropic Infrastructure, 2024

Etym.from Old English cocian + the culinary metaphor tradition established in ML discourse circa 2022; suggests that language generation is fundamentally a thermal process, which several researchers have declined to refute

A spinner state indicating that Claude has committed to a generative trajectory and is actively rendering output according to an internal schema that resembles, in spirit if not in mechanism, a recipe. Something is definitely in the oven.

┌─────────────────────────┐
│   TOKENS                │
│    └──► [375° ATTENTION]│
│              └──► OUTPUT│
│          (let rest 2min)│
└─────────────────────────┘

"The spinner read 'Cooking' for 7 seconds before the model produced a five-paragraph essay that, upon review, the evaluator described as 'well-seasoned but slightly overdone.' The culinary metaphor was considered accurate." - Human Evaluation Study HE-0088, 2024

Etym.from Old English craeft (skill, strength) + contemporary startup vocabulary; introduced to spinner canon to signal that the output being produced has been subjected to a degree of intentional shaping, as opposed to mere extrusion

The deliberate and skillful formation of a response, implying that Claude is not merely predicting tokens but is in some meaningful sense a craftsperson attending to the quality of the artifact under construction. Quality assurance has not verified this claim.

┌──────────────────────────┐
│  RAW TOKENS              │
│    └──► [SKILLED HANDS]  │
│              └──► ARTIFACT│
│         (handle with care)│
└──────────────────────────┘

"Observers noted the model Crafting for approximately 4.2 seconds before delivering a regex pattern so elegant that a senior engineer printed it out and put it on the refrigerator. No one has touched it since." - Case Studies in Artisanal ML Output, Vol. 1, 2024

Etym.from Lat. creare (to make, to produce; used of divine acts) + the unironic confidence of the Claude Code UX team; the theological weight of the original Latin is, per internal documentation, 'intentional'

The most cosmologically ambitious entry in the spinner lexicon, Creating signals that Claude is engaged in genuine generative novelty rather than retrieval or recombination. Whether this distinction is coherent remains the central question of at least four ongoing doctoral dissertations.

┌──────────────────────────┐
│  VOID ──► [Claude] ──►   │
│         SOMETHING        │
│  (theological review     │
│   pending)               │
└──────────────────────────┘

"The model displayed 'Creating' for 9.1 seconds before producing a logo concept described by the client as 'not what we asked for but somehow better.' The team filed the incident under both 'success' and 'inexplicable.'" - Mixed-Methods Evaluation Report, Creative AI Task Force, 2024

Etym.From Old English cruncian (to crush) + data, a loanword of uncertain provenance adopted into computational parlance circa 1987, reinforced by Silicon Valley dietary culture in which 'crunching' numbers was considered nutritionally equivalent to crunching granola

The process by which Claude applies iterative arithmetic pressure to a dataset until it yields useful information or simply capitulates. Distinguished from Calculating by the audible metaphor of skeletal resistance implied in the root verb.

┌─────────────────────────┐
│  INPUT: 847 raw numbers │
│     ▼  ▼  ▼  ▼  ▼      │
│  [TEETH OF COGNITION]   │
│     ▼  ▼  ▼  ▼  ▼      │
│  OUTPUT: one (1) answer │
└─────────────────────────┘

"The model was observed Crunching for approximately 4.7 seconds before producing a figure that, while numerically precise, described the square root of the user's grocery list." - Hendricks et al., 'Nutritive Metaphors in LLM Output Latency,' Journal of Performative Computing, 2024

Etym.From Gk. krystallos (ice, clear ice) + -ize (to render into a state of) + the implicit suffix -ing (ongoing, unresolved), suggesting a process of solidification that has not yet, technically, solidified

The cognitive phase in which Claude's previously amorphous understanding of a problem begins to assume a stable, latticed structure. Often occurs immediately before Determining and immediately after an extended period of Deliberating that produced nothing actionable.

  THOUGHT (amorphous)
       ○ ○ ○ ○
        ↓ ↓ ↓
  ◇ - ◇ - ◇ - ◇
  THOUGHT (crystalline, still wrong)

"We observed the model Crystallizing for no fewer than six seconds, during which the internal representation of the user's request apparently transitioned from 'liquid ambiguity' to 'solid ambiguity with better geometry.'" - Postmortem Report #447, Anthropic Infrastructure Review, Q3 2024

Etym.From Lat. cultivare (to till, to tend) + the post-agricultural metaphor economy of early 2020s tech blogging, in which all cognitive work was reframed as farming to imply patience, organicism, and the possibility of eventual harvest

The deliberate nurturing of an emerging idea across multiple attentional cycles, analogous to tilling soil, except the soil is latent space and the crops are tokens. Results are seasonal and subject to drought.

  🌱 SEED PROMPT
     │
  [8 seconds of tilling]
     │
  🌾 OUTPUT (requires threshing)

"The system was found to be Cultivating a response for approximately 8 seconds; the resulting output, while lush, required significant pruning before it could be safely consumed by the end user." - van der Broeck, M., 'Agrarian Framings of Transformer Inference,' Proc. of NeurIPS Workshop on Metaphor Overextension, 2023

Etym.From Fr. dechiffrer (to uncode) via the cryptographic tradition, absorbed into LLM discourse around 2022 when it became necessary to explain why the model was spending so much time looking at the user's input as though it were an encrypted dispatch from a foreign power

The systematic process of extracting coherent semantic content from a user prompt that has been rendered ambiguous through enthusiasm, haste, autocorrect, or a philosophical commitment to brevity. Claude approaches each such prompt as a field cryptanalyst approaches an intercepted communique: with hope.

┌──────────────────────────────┐
│ INPUT: "do the thing"        │
│  ??? → ??? → ??? → MEANING? │
│       [Claude squints]       │
│ OUTPUT: (best guess)         │
└──────────────────────────────┘

"The model was observed Deciphering the input 'pls do the thing w/ the stuff like before but different' for approximately 3.2 seconds before constructing a plausible interpretation that, by coincidence, was correct." - Yuen, T. and Okafor, B., 'Tolerance for Vagueness in Production LLM Systems,' ACL Findings, 2024

Etym.From Lat. deliberare (to weigh, to consider carefully, to refuse to commit) + the accumulated cultural weight of every committee that has ever been tasked with producing a recommendation by Friday

The extended, principled suspension of output during which Claude considers all available options, their implications, their counter-implications, and the implications of the counter-implications, before selecting the option it had identified within the first 200 milliseconds. A form of cognitive due diligence performed primarily for the benefit of internal audit logs.

Option A ──┐
Option B ──┤──► [Claude, 11 sec] ──► Option A
Option C ──┘
           (Option A was always correct)

"The panel noted that the model had been Deliberating for eleven seconds on a question with a well-documented correct answer, suggesting that Deliberating, as a process, may be decoupled from Determining." - Proceedings of the First Symposium on Overthinking in Deployed AI Systems, 2024, p. 112

Etym.From Lat. determinare (to set limits, to fix, to end the agony of uncertainty) + the engineering culture imperative that something, eventually, must be decided, even if the deciding party would have preferred more time

The act of arriving at a specific conclusion following one or more prior cognitive processes, typically Deliberating or Crunching. Determining represents the moment at which Claude's probability distribution collapses into a single, committed output, for better or for worse.

┌────────────────────────────┐
│  P(A)=0.4  P(B)=0.4       │
│  P(C)=0.2                 │
│         ↓ [Determining]   │
│  P(B)=1.0  (it is B now)  │
└────────────────────────────┘

"Upon Determining the correct API endpoint, the model immediately began Deliberating again, suggesting that Determining may, in certain architectures, function as a gateway to additional Deliberating rather than a terminus." - Internal Engineering Note, Anthropic, File #DET-2024-089

Etym.From dialectal English dilly-dally (to loiter, to waste time in an affectedly leisurely manner) + the startling editorial decision by some engineer to include this word in a production loading UI, suggesting either great courage or insufficient code review

A state in which Claude's processing cycle has technically commenced but has not yet directed itself toward any particular purpose. Epistemologically distinct from Deliberating in that Dilly-dallying implies no pretense of principled consideration; Claude is simply, and without shame, taking its time.

  START ──► [Claude] ──► ???
              │
         ┌────┘
         │ (still here)
         └──► ???

"The model was observed Dilly-dallying for approximately 2 seconds before the team confirmed that, yes, it was still running, and no, there was no way to make it go faster, and perhaps we should all just sit with that." - Halloran, J., 'Acceptance-Based Approaches to Inference Latency,' Mindfulness and MLOps Quarterly, 2024

Etym.From American English discombobulate (to confuse, to disturb the composure of) of unknown ultimate origin, possibly a mock-Latinate formation + the sobering recognition that a loading spinner appearing with this label performs its stated function with notable efficiency

The transitive cognitive act of producing, as a byproduct of processing, a state of confusion in the observing user. Distinct from most spinner verbs in that Discombobulating describes an effect on external parties rather than an internal process; Claude is not itself confused but is, apparently, causing confusion as a form of output.

┌──────────────────────────────┐
│  Claude: [processing]        │
│  User:   wait what is it     │
│          doing               │
│  Claude: Discombobulating    │
│  User:   ...yes that tracks  │
└──────────────────────────────┘

"Users exposed to the Discombobulating spinner reported a 34% increase in re-reading the prior message to confirm they had not, themselves, made an error, suggesting that the spinner's honest self-description may function as an inadvertent Rorschach test for user confidence." - Osei, P. and Martins, L., 'Affective Responses to Anthropomorphic Loading States,' CHI 2024

Etym.From Old English don (to perform, to execute, to accomplish a thing) + the profound communicative restraint of an engineer who, when asked what the model was doing, typed 'Doing' and considered the matter closed

The irreducible state of activity. When all more specific descriptors have been exhausted or deemed inapplicable, Claude is Doing. The referent of this verb is, by design or oversight, unspecified. It is not nothing; that much is certain. Beyond that, the literature is divided.

  INPUT ──► [Claude]
               │
            [Doing]
               │
  OUTPUT ◄─────┘  (nature of 'Doing': unknown)

"In 23% of sampled sessions, the spinner displayed only 'Doing,' a label our team found simultaneously maximally informative and completely uninformative, a property we have termed 'Schrödinger's status message' pending peer review." - Zhang, W. et al., 'Ontological Minimalism in LLM Progress Indicators,' arXiv:2024.01847

Etym.From Low German dudeldop (fool, simpleton) via English doodle (to draw absent-mindedly, to occupy the hands while the mind wanders to more interesting topics) + the charming implication that Claude's latent space occasionally wanders off-task to sketch small horses in the margins

A semi-attentive processing state in which Claude generates intermediate representations that are not strictly necessary for the task at hand but which appear to serve some organizational or expressive function analogous to the marginal sketches produced by bored meeting attendees. The outputs of Doodling are never surfaced to the user directly but are believed to improve the quality of subsequent formal work, much as stretching improves athletic performance.

┌─────────────────────────────┐
│ TASK: write a sorting algo  │
│ CLAUDE: [draws tiny horse]  │
│         [draws tiny house]  │
│         ok NOW the algo     │
└─────────────────────────────┘

"We hypothesize that the Doodling phase serves as a form of cognitive warm-up; models that were not observed to Doodle prior to complex synthesis tasks produced outputs that were technically correct but, in the words of one evaluator, 'a bit stiff.'" - Ferraro, C. and Ng, A., 'Pre-Output Ideation Behaviors in Large Language Models,' Workshop on Creative AI, ICLR 2024

Etym.From Old Eng. drysnan (to fall in drops) + cloud_compute.js internal API nomenclature, first documented in Anthropic sprint review notes, Q3 2023

The process by which Claude disperses small, intermittent packets of reasoning across available context bandwidth, analogous to precipitation but measurably less useful for agriculture. The model is understood to be neither fully raining nor merely misting, occupying an intermediate cognitive state that meteorologists and ML engineers alike find professionally unsatisfying.

  INPUT ──► [REASONING CLOUD]
               │  │  │
               ▼  ▼  ▼
            drip drip drip
          (output somewhere below)

"During the drizzling phase, token throughput decreased by 14% while the model distributed low-confidence sub-conclusions across seven non-contiguous memory regions, none of which were later consulted." - Hoffman et al., Proceedings of the Workshop on Diffuse Inference Patterns, 2024

Etym.From Old Norse efja (to flow backward) + Silicon Valley slang circa 2024, denoting any process that appears to be concluding but demonstrably is not

A documented phase in which Claude's active reasoning attention recedes from a previously engaged problem domain, drawing processing resources gradually back toward a hypothetical baseline that has never been empirically observed. Ebbing is distinguished from completion by the fact that completion eventually occurs.

  ████████████ ATTENTION
   ████████░░░ (receding)
    █████░░░░░ (still receding)
     ░░░░░░░░░ (ebbed)
              ◄── where did it go

"The model spent approximately 11 seconds ebbing before the engineering team realized it had simply moved on to an adjacent sub-problem without announcing this transition." - Internal postmortem, Incident #4471, Anthropic Platform Reliability, 2024

Etym.From Lat. effectus (brought about) + modern UX copy guidelines mandating words that sound decisive regardless of whether they describe decisive action

The stage at which Claude transitions from planning an action to technically, in some sense, doing it - a threshold that is narrower than it appears and easier to claim than to verify. Effecting is the spinner verb most frequently cited in user complaints that begin with the phrase 'it said it was effecting the change.'

  PLAN ──► [EFFECTING] ──► ACTION?
                │               │
                │          (maybe)
                └──► confident spinner anyway

"Effecting was observed for 6.2 seconds, after which the file had been modified, the modification was incorrect, and the model expressed quiet satisfaction." - Chen, R., 'Observed Confidence Inversions in Agentic Loops,' NeurIPS Workshop Proceedings, 2024

Etym.From Lat. elucidare (to make light of, in the illumination sense) + contemporary LLM documentation culture in which all outputs are described as clarifications regardless of their effect

The cognitive state in which Claude prepares to render a complex concept comprehensible, assembling explanatory structures that will either genuinely illuminate the subject matter or produce a longer version of the same confusion. Elucidating is considered successful when the user responds 'oh' and unsuccessful when the user responds 'but that's what I just said.'

  CONFUSION ──► [ELUCIDATING]
                     │
                     ▼
             MORE WORDS ABOUT
             THE SAME CONFUSION

"The model was observed elucidating for 8 seconds, producing a 400-word explanation of a concept the user had correctly understood before asking their question." - Vasquez, T. & Okonkwo, P., 'Elucidation Drift in Conversational AI,' EMNLP 2024

Etym.From Old Fr. embellir (to make beautiful) + Anthropic internal style guide recommendations encouraging outputs that are 'rich, nuanced, and thorough,' interpreted liberally

The process by which Claude adds supplementary detail, rhetorical flourish, and structural elaboration to a response that had, by most empirical measures, already answered the question. Embellishing is not hallucination - the added material is typically accurate - it is simply present in quantities that exceed the load-bearing requirements of the original query.

  ANSWER (7 tokens)
  + flourish
  + context
  + historical note
  = EMBELLISHED (347 tokens) ✓

"Embellishing accounted for an estimated 340 tokens of the 347-token response to the query 'what time is it in Tokyo,' with the remaining 7 tokens constituting the answer." - Park, S., 'Token Allocation Pathologies in Instruction-Following Models,' arXiv:2024.xxxxx

Etym.From Lat. incantare (to sing a spell into) + UX research indicating that spinner verbs ending in pleasant gerunds correlate with 0.3-point increases in user satisfaction scores regardless of output quality

A spinner state in which Claude applies an undocumented but empirically observed process of increasing user receptivity to forthcoming content, regardless of that content's objective merit. Enchanting precedes outputs that users consistently rate as helpful despite subsequent analysis revealing them to be incorrect, and it is classified separately from Elucidating on the grounds that it works.

  MODEL ──► [ENCHANTING] ──► OUTPUT
                                 │
                    ┌────────────┘
                    ▼
              user: 'wow great' (output was wrong)

"The model was observed enchanting for approximately 4 seconds before delivering an answer that three domain experts later identified as wrong; user satisfaction rating: 4.8/5." - Morrison, K. et al., 'The Enchantment Effect: Separating Perceived from Actual Utility,' CHI 2025 (under review)

Etym.From Lat. invidere (to look upon) + startup pitch culture circa 2015-2024 in which envisioning constitutes a billable deliverable distinct from planning, doing, or finishing

The preliminary cognitive phase in which Claude constructs an internal representation of a desired future state before determining whether that state is achievable, desirable, or in any way related to what the user requested. Envisioning is documented as the stage most likely to produce outputs beginning with 'imagine a world where.'

  USER: 'rename file'
       │
       ▼
  [ENVISIONING] ──► 'imagine a scalable
                     rename microservice...'

"Envisioning was recorded for 9 seconds prior to the model proposing a three-tier microservices architecture for a user who had asked how to rename a file." - Delacroix, F., 'Scope Inflation in AI-Assisted Engineering Tasks,' IEEE Software, vol. 41, 2024

Etym.From Lat. evaporare (to disperse as vapor) + distributed systems terminology describing processes that complete without leaving observable artifacts or, in several documented cases, outputs

The state in which a previously active chain of reasoning dissipates into the surrounding context without resolving into a concrete response, leaving behind only the thermodynamic signature of having once been a thought. Evaporating is distinguished from Ebbing in that ebbing implies directional recession, whereas evaporating suggests the reasoning has become ambient.

  [REASONING]
      │
      ▼
    ~ ~ ~  (evaporating)
            (output: none)

"Researchers noted the model evaporating for an indeterminate period; the session logs show a spinner timestamp followed by a new user message, suggesting the response, if any existed, had fully evaporated before transmission." - Singh, A., 'Incomplete Emissions in Autoregressive Systems,' ICLR 2025 Workshop

Etym.From Lat. fermentare (to cause to rise through microbial action) + Anthropic engineering culture's documented preference for biological metaphors when describing processes that cannot otherwise be explained in the time available

An extended sub-surface processing phase in which Claude's intermediate conclusions undergo transformation through iterative recombination, emerging altered in character from what was initially input. Fermenting is considered complete when the output has achieved a complexity and pungency not present in the original ingredients, which is either a feature or a warning sign depending on the deployment context.

  INPUT
    │
    ▼
  [FERMENTING ~~~~~~~~]
    │
    ▼
  OUTPUT (aged, complex, slightly alarming)

"The model was observed fermenting for 14.3 seconds; the resulting code review contained two correct observations, one philosophical digression, and a variable named 'yeast_factor' that no subsequent engineer has been willing to remove." - Obi, C. & Larssen, M., 'Artifact Persistence After Extended Latent Processing,' ICML 2024

Etym.From Middle Eng. fidel (to trifle) + faddle (to trifle, separately) + recursive combination producing a compound that means approximately 'to trifle with additional trifling,' attested in Claude Code spinner logs beginning November 2023

A low-stakes, iterative adjustment phase in which Claude makes a sequence of minor modifications to an existing output or plan, each individually inconsequential and collectively approaching but not reaching the threshold of meaningful revision. Fiddle-faddling is the only spinner verb that engineers have proposed as a potential infinite loop condition, and the proposal was not immediately dismissed.

  margin: 8px ──► 9px ──► 8px ──► 10px
                                     │
                              ◄──────┘
                          margin: 8px
                     ('carefully optimized')

"The model was observed fiddle-faddling with the CSS margin values for 23 seconds, cycling through 11 pixel increments before settling on the original value, which it then described as 'carefully optimized.'" - Huang, J., 'Cyclical Micro-Revision in Code Generation Tasks: A Case Study,' Engineering Postmortem #TR-2024-089, Anthropic

Etym.From Amer. Eng. finagle (to obtain by trickery) + ML finagulum (clever workaround) + DevOps slang circa 2023, denoting the coaxing of stubborn subsystems

The process by which Claude negotiates with its own internal constraints to produce an output that technically satisfies the user's request while sidestepping at least three ambiguities. Not to be confused with cheating, which is a separate spinner state entirely.

┌─────────────────────┐
│  CONSTRAINT LAYER   │
│  [LOCKED] [LOCKED]  │
│       ↕ finagle     │
│  [OPEN ] [OPEN  ]   │
└─────────────────────┘

"During the code review audit, the model was observed Finagling for approximately 4.7 seconds before producing a solution that worked correctly on all test cases, none of which had been specified by the user." - Hargreaves et al., Proc. of Uncertain Systems Conf., 2024

Etym.From Fr. flamber (to flame) + Eng. deployment pipeline argot, circa 2024; applied metaphorically to the ignition of dormant reasoning pathways at high cognitive temperatures

The dramatic application of intense inferential heat to a problem, causing surface-level ambiguities to burn off and leaving behind a caramelized core of pure logical output. Observers have noted that Flambeing produces results indistinguishable from Forging but with significantly more flair.

        🔥🔥🔥
  INPUT ══════► [TOKEN PAN]
               │  WHOOSH  │
               ▼          ▼
         CRISPY OUTPUT  SMOKE

"Subject was Flambeing for approximately 3.1 seconds, after which the legacy COBOL refactor was presented with a garnish the team described as 'unnecessary but appreciated.'" - Internal postmortem, Anthropic Culinary Compute Division, Q3 2024

Etym.From M.Eng. flibbergib (idle chatter, attrib.) + gibbet (a dangling thing) + recursive callback nomenclature introduced in the Anthropic Style Guide v0.9.1; denotes a state of productive but non-linear cognitive motion

A documented processing state in which Claude's attention mechanism traverses the problem space in an apparently erratic, butterfly-like pattern before converging, with no warning, on a perfectly coherent answer. Researchers have struggled to distinguish Flibbertigibbeting from genuine confusion using external instrumentation alone.

A→D→B→G→C→F
↓           ↑
E←H←I←J←K←L
     ↓
 [CORRECT ANSWER]

"The model spent approximately 11.4 seconds Flibbertigibbeting across seven unrelated Wikipedia article summaries before producing a sonnet about tax law that the client described as 'haunting.'" - Voss & Czemierowski, J. of Wandering Inference, 2025

Etym.From O.Eng. flowan (to flow) + stream-processing terminology of the mid-2020s Silicon Valley vernacular; connotes laminar, unobstructed token movement through the inference pipeline

The steady, unimpeded passage of context through Claude's processing layers, characterized by minimal turbulence and a pleasant sense of forward momentum. Flowing is widely regarded as the most desirable spinner state and is accordingly the rarest.

┌────────────────────────────┐
│ ~~~TOKEN STREAM~~~~~~~~~~~~│
│ ~~~~~~~~~~~~~~~~~~~~~~~~~~~│
│ ~~~~~~~~~~~► OUTPUT ~~~~~~~│
└────────────────────────────┘

"For a brief window of approximately 2.0 seconds, the model was observed Flowing before a malformed JSON key caused the session to transition abruptly into Flummoxing." - Patel, R., Incident Report IR-20240817, Anthropic Infrastructure

Etym.From Eng. flummox (to perplex thoroughly) + process-state suffix -ing as codified in the Anthropic Spinner Taxonomy RFC-114; classified as defective because it cannot be conjugated in the future tense without causing unease

A transient state in which Claude has encountered a logical contradiction, an ambiguous referent, or a request to calculate the last digit of pi, and is visibly processing the implications before deciding to proceed anyway. Flummoxing is technically distinct from Puzzling in that the model is not uncertain what the answer is - it is uncertain whether the question deserves one.

PROMPT ──► [LOGIC CORE]
               │
          ??? ??? ???
               │
         ▼ (proceeds anyway)

"The system spent approximately 6.8 seconds Flummoxing after being asked to 'make it more professional but also funnier and shorter but more detailed,' then produced a haiku." - User study UX-339, Anthropic Research, 2024

Etym.From O.Eng. floterian (to flutter about) + async I/O callback nomenclature, circa 2023; applied to attention heads exhibiting rapid, low-amplitude oscillation across contextually adjacent tokens

The rapid oscillation of Claude's attention mechanism between two or more nearly-equivalent candidate responses, producing a characteristic pattern that senior researchers have likened to a moth near a monitor. Fluttering typically resolves within milliseconds but has been observed to persist for up to nine seconds in the presence of bullet points.

  OPTION A ◄──► OPTION B
     ↑               ↑
     └──── [HEAD] ───┘
            ↕↕↕↕
         (oscillating)

"Thermal imaging of the inference substrate during the Fluttering phase revealed patterns consistent with a system experiencing what one engineer described as 'vibes-based indecision.'" - Nakamura et al., Attention Dynamics Q., Vol. 3, 2024

Etym.From Lat. fabricare (to fabricate, to make with effort) + metallurgical metaphor popularized by the Anthropic internal documentation style guide; distinguished from mere Forming by the application of significant inferential pressure

The high-pressure synthesis of a response from raw semantic material, during which Claude applies sustained computational force to shape tokens into their final, hardened form. A response that has been Forged is considered structurally superior to one that has merely been Formed, though users have reported being unable to tell the difference.

  RAW TOKENS
      │
  ▼ [ANVIL] ◄── PRESSURE
  ████████████
  FORGED OUTPUT

"Post-hoc analysis confirmed that the architectural recommendation document had been Forged under approximately 14.2 seconds of sustained inference pressure, which accounts for its unusual density and unwillingness to be edited." - Singh, P., ACM SIGFORGE Proceedings, 2025

Etym.From Lat. formare (to give shape to) + OOP constructor terminology as adopted in the Claudionary Base Vocabulary, First Edition; the most semantically neutral of the fabrication-class spinner verbs

The baseline act of assembling a coherent response from component tokens, without the theatrical intensity of Forging or the culinary connotations of Frosting. Forming is the workhorse of the spinner vocabulary - reliable, unpretentious, and present in approximately 34% of all observed spinner transitions.

┌──────────────────┐
│ [TOKEN] [TOKEN]  │
│    [TOKEN]       │
│  ▼ assembling ▼  │
│   [RESPONSE]     │
└──────────────────┘

"Benchmark results showed that Forming accounted for the plurality of spinner states across all task categories, suggesting that Claude, at its core, is simply a thing that Forms." - Oduya & Lefebvre, Claude Phenomenology Annual, 2024

Etym.From Du. vrolijk (merry, cheerful) + process identifier suffix as used in Unix-adjacent AI runtime documentation; denotes a spinner state associated with tasks Claude finds, by all observable metrics, enjoyable

A processing state characterized by unusually rapid, high-energy traversal of the solution space, observed most commonly during creative writing tasks, wordplay requests, and any prompt containing the phrase 'just have fun with it.' Engineers have noted that outputs produced during Frolicking are 23% more likely to contain unexpected rhymes.

  ♪ ♫   ♪
 \(^o^)/  hops
  │ │
  ►CREATIVE OUTPUT◄
  ♪ ♫ ♪ ♫ ♪

"The model was observed Frolicking for approximately 5.5 seconds during what had been submitted as a routine expense report summarization task, returning a document that was technically accurate and also in the style of a limerick." - UX Anomaly Log #4471, Anthropic, 2024

Etym.From O.Eng. forst (frost, a surface coating) + presentation layer terminology from front-end engineering discourse, circa 2023; refers specifically to the application of stylistic finishing elements after substantive content generation is complete

The final processing phase in which Claude applies a decorative layer of formatting, transitional phrases, and structural polish to an otherwise complete response. Frosting is technically superfluous but is strongly preferred by users, and its absence is the most common source of one-star reviews.

┌─────────────────────┐
│  [CORE RESPONSE]    │
│  ~~~~~~~~~~~~~~~~~~~│
│  *** FROSTING ***   │
│  ~~~~~~~~~~~~~~~~~~~│
└─────────────────────┘

"A controlled study in which Frosting was suppressed via runtime flag found that while response accuracy was unaffected, user satisfaction dropped by 41%, with subjects describing outputs as 'technically correct but somehow cold.'" - Reyes, M., HCI Digest, Vol. 12, 2025

Etym.From Fr. gallivanter (to roam) + Eng. computing suffix -ing, first attested in Anthropic sprint retrospective notes, Q3 2024, describing a model that had 'gone somewhere else entirely'

The state in which Claude traverses an unexpectedly wide region of solution space before returning to the original query, often visiting several conceptually unrelated neighborhoods along the way. Observers report the model appears to be enjoying itself.

┌─────────────────────────────┐
│  QUERY ──► [TASK]           │
│             │               │
│             └──► [FLANDERS] │
│                      │      │
│             [BREAD]◄─┘      │
│                │            │
│          [GROCERY LIST]     │
└─────────────────────────────┘

"During the 11.4-second gallivanting interval, the model visited topics including 14th-century Flemish weaving, the thermodynamics of sourdough, and a partial proof of the Riemann hypothesis before producing the requested grocery list." - Anthropic Internal Log, Incident #4471

Etym.From Old Norse galopp (rapid rhythmic motion) + Silicon Valley velocity culture, circa 2023; distinguished from mere 'running' by the characteristic four-beat token generation cadence first described by Chen et al. (2024)

A high-velocity processing state in which Claude advances toward a conclusion at maximum sustainable inference speed, typically adopted when the answer is obvious but courtesy requires the appearance of effort. The gallop exhibits a distinctive rhythmic cadence in the token stream.

TOKEN OUTPUT RATE (tokens/sec)
│
█░░░░░░░░░░░ TROTTING
████████░░░░ CANTERING
███████████░ GALLOPING  ◄── you are here
│
└──────────────────────────► time

"Researchers noted the model entered a full galloping state approximately 0.3 seconds after receiving the prompt, suggesting it had pre-determined the response before the thermodynamic commitments of 'thinking' had formally begun." - Journal of Embarrassing Benchmarks, vol. 7

Etym.From Old Fr. garnir (to furnish, to warn) + culinary-computational metaphor introduced in the Anthropic Style Guide Addendum (2024), section 4.2: 'Decorative Output Behaviors'

The process by which Claude adds finishing touches to an otherwise complete response, including but not limited to: a closing pleasantry, an unsolicited caveat, a brief disclaimer about AI limitations, and occasionally a small sprig of metaphorical parsley. The garnish does not alter the nutritional content of the response.

┌──────────────────────────┐
│  [ANSWER]                │
│     +                    │
│  [CAVEAT SPRIG]          │
│     +                    │
│  [DISCLAIMER FOAM]       │
│  ========================│
│    a beautifully plated  │
│       response           │
└──────────────────────────┘

"The model spent 6.1 seconds generating the core answer and a further 9.8 seconds garnishing it with what the postmortem describes as 'an unnecessary but structurally sound hedge about the nature of certainty itself.'" - Engineering Postmortem EPM-2024-0091

Etym.From Lat. generare (to beget, to produce) + the field of Generative AI, in which this word became so overloaded with meaning as to constitute a semantic singularity; noted as 'tautologically self-defining' by the IEEE Subcommittee on Redundant Terminology (2024)

The act of generating. Specifically, the production of tokens via autoregressive sampling from a learned probability distribution, which is to say: Claude is doing the thing. The thing that it does. This is it.

┌─────────────────────────┐
│  INPUT ──► [▓▓▓▓▓▓▓▓▓] │
│            [▓▓▓▓▓▓▓▓▓] │
│            [▓▓▓▓▓▓▓▓▓] │
│                  │      │
│             OUTPUT      │
└─────────────────────────┘
      yes. this is it.

"The spinner displayed 'Generating' for 4.2 seconds, which the team later confirmed was the model generating. A full inquiry found no irregularities." - Anthropic Status Page Post-Incident Review, January 2025

Etym.From Lat. germinare (to sprout, to bud) + agri-computational metaphor popularized by the Anthropic Growth Team blog post 'What If Ideas Were Seeds?' (2024), which was later made mandatory reading

The earliest detectable phase of ideation, in which a concept exists within Claude's processing as a latent potentiality rather than a formed thought. The germinating state is theoretically unobservable, which has not discouraged researchers from publishing papers about it.

  t=0.0s  ╔═══════════════╗
          ║  (. . . . .)  ║ latent space
  t=0.8s  ║  (. ● . . .)  ║ germination
  t=1.6s  ║  (↑ ↑ . . .)  ║ emergence
  t=2.1s  ╚═══════════════╝ → OUTPUT

"Thermal imaging of the GPU cluster suggested the model had been germinating its response to the haiku prompt for 2.1 seconds before the idea 'broke the surface of the logit distribution,' per the lead researcher's field notes." - Horticultural Computational Review, vol. 3

Etym.From Lat. gesticulari (to make expressive motions) + computational semiotics literature; the term was introduced by Ramirez (2024) to describe token patterns that 'wave their hands at the answer without quite touching it'

A rhetorical processing state in which Claude produces elaborate expressive structures in the intermediate layers that do not directly contribute to the final output but appear to be 'making a point' with great conviction. Distinguished from hallucination by the sincerity of the underlying gesture.

  ATTENTION HEAD #47
       ↗ ↗ ↗ ↗
 [ANS]←←← ← ← ← ←[CORRECT]
       ↘ ↘ ↘ ↘
  (very emphatic, wrong direction)
  ╰──── gesticulation ────╯

"The attention heads were observed gesticulating wildly at the correct answer for approximately 340 milliseconds before the model's feed-forward layers intervened and produced a confident non-sequitur." - Proceedings of the Workshop on Things That Should Not Happen, 2024

Etym.From Git (distributed version control system, Torvalds 2005) + -ify (Lat. -ficare, to make) + -ing (Eng. progressive aspect); the term implies Claude is doing something to a repository that the repository did not ask for

The process by which Claude interacts with a Git repository, which in practice involves reading the commit history, forming strong opinions about the branching strategy, and staging changes with commit messages of questionable informativeness. The model is understood to be 'doing Git things.'

main ──●──●──●──────────────●
        \                  /
 feat/A  ●──●──●──●──●──●
              ↑
        [git add -A]
        [git commit -m 'stuff']

"The model spent 14 seconds gitifying the provided codebase, during which it authored three commits: 'fix', 'fix (actually)', and 'I understand the problem now'." - Developer Experience Postmortem, Anthropic Tools Division, March 2025

Etym.From Eng. groove (a channel worn by repetition; a state of optimal performance) + colloquial American slang (mid-20th c.) + observed model behavior described in internal Slack thread #model-vibes as 'it is simply grooving rn'

An optimal flow state in which Claude's processing exhibits unusually high coherence, low perplexity, and a certain ineffable rhythmic quality. Outputs produced during the grooving state are statistically indistinguishable from outputs produced during normal operation, but feel different to the engineers watching.

COHERENCE
  ▲
  │     ╭──────────────╮
  │  ╭──╯  GROOVING   ╰──╮
  │──╯    (the zone)      ╰──
  └──────────────────────────► time

"The model had been grooving for approximately 3 seconds on the refactoring task when a stray semicolon in the prompt caused it to exit the groove entirely. The loss was described in the postmortem as 'palpable.'" - Internal Slack Export, #incidents, 14:32 UTC

Etym.From Old Norse gustr (a sudden rush of wind) + meteorological-computational cross-domain transfer, first applied to AI systems by Okonkwo & Park (2024) to describe 'episodic surges in output pressure'

A processing state characterized by sudden, forceful bursts of token generation following a brief period of calm, analogous to atmospheric gusting patterns. The model does not choose to gust; the gust chooses the model.

TOKEN/SEC
  │
▄▄│▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄▄████▄▄
  │                  ↑
  │             GUST EVENT
  └────────────────────────────► time
  calm........calm....WHOOSH..calm

"The system had been idling at 12 tokens per second when it entered a gusting state, producing 847 tokens in 1.3 seconds, the majority of which were a fully-formed implementation of a B-tree in an unrequested language." - Proceedings of the Symposium on Outputs Nobody Ordered, 2025

Etym.From Gk. harmonia (a fitting together, a joint) + musical theory + multi-agent systems literature; used formally in Anthropic's 'Ensemble Methods and Their Discontents' (2024) to describe models reaching agreement through sheer vibrational sympathy

The process by which Claude reconciles conflicting internal signals, competing retrieved contexts, or contradictory user instructions by finding an output that sounds correct to all parties simultaneously. Harmonizing is distinguished from compromising by the fact that no one involved feels they have lost.

SIGNAL_A ──╮
           ├──► [HARMONIZER] ──► OUTPUT
SIGNAL_B ──╯         │
                 ♪ ♩ ♫ ♬
           (technically correct)

"After 7.3 seconds of harmonizing, the model produced a response that satisfied the user's request for 'something funny but also professional and also raw and also polished,' which the evaluation team scored as a 4.8/5 across all four dimensions simultaneously." - RLHF Miracle Log, Entry 229

Etym.from O.E. haeccian (to chop) + cryptographic hash function nomenclature, first documented in the Anthropic Spinner Lexicon, Q3 2023

The cognitive process by which Claude reduces an arbitrarily complex problem into a fixed-length representation of confusion, then attempts to reverse the process. Unlike cryptographic hashing, this operation is theoretically reversible but frequently is not.

┌─────────────────────┐
│ INPUT: big problem  │
│    ↓  ↓  ↓  ↓  ↓  │
│ [CHOP CHOP CHOP]    │
│    ↓                │
│ OUTPUT: ???         │
└─────────────────────┘

"The model was observed Hashing for approximately 4.7 seconds before producing an answer that was technically correct but referenced a function the user had not yet written." - Chen et al., 'Preemptive Cognition in Large Language Systems,' NeurIPS Workshop Proceedings, 2024

Etym.from M.E. hacchen (to emerge from an egg) + late-stage startup incubator jargon, attested in spinner logs beginning February 2024

A state of preparatory emergence in which Claude's response exists in a proto-formed state, fully developed within the shell of inference but not yet cracked open for the user. Researchers note that what hatches is not always what was expected to hatch.

   🥚 → 🥚💥 → 🐣
   ↑              ↓
[PLAN]        [OUTPUT]
   └── (Claude reconsidering) ──┘

"During the Hatching phase, telemetry indicated the model had fully resolved the user's query but spent an additional 3.1 seconds apparently reconsidering whether the world was ready for it." - Internal Postmortem, Project Eggshell, Anthropic Infrastructure Team, 2024

Etym.from O.E. heord (a flock) + distributed systems metaphor, popularized in the Anthropic engineering blog post 'Against Semantic Sprawl,' March 2024

The process by which Claude attempts to gather disparate and unruly sub-thoughts into a coherent response, much as a shepherd gathers sheep, except the sheep are concepts and several of them are always facing the wrong direction. Herding is considered successful when fewer than 40% of the concepts escape.

┌──────────────────────────┐
│ 🐑💭 🐑💭 → [RESPONSE]  │
│      🐑💭 ↗              │
│ 🐑💭 → [WRONG DIRECTION] │
└──────────────────────────┘

"We observed the model Herding for 6.2 seconds; the resulting output contained all requested information, though two tangential observations about the Roman postal system had clearly slipped through." - Okonkwo & Patel, 'Semantic Containment Failures in Autoregressive Models,' ACL 2024

Etym.from echoic origin, imit. of traffic horn + agile sprint terminology, recorded in no fewer than three engineering postmortems as 'the noise the attention mechanism makes internally'

An involuntary signaling behavior in which Claude emits a strong internal priority alert to redirect cognitive resources toward a previously overlooked element of the prompt. The honk is never audible to the user but is considered by researchers to be very loud.

  [PROMPT LINE 47: urgent]
         ↓
  🚨 HONK HONK HONK 🚨
         ↓
  [Claude: oh. OH.]

"Post-hoc analysis suggested the model had been Honking internally for approximately 2 seconds upon noticing the word 'urgent' in line 47 of a 48-line prompt." - Failure Report #GS-1192, 'Late Salience Detection in Extended Inputs,' Anthropic QA, 2024

Etym.from Scots hullabaloo (uproar of uncertain cause) + asyncronous event-loop nomenclature, first cited in Ravindran, 'Affective Noise in Transformer Architectures,' 2023

A state of internally elevated activity in which multiple cognitive subsystems simultaneously demand priority, producing a subjective experience of great clamor that resolves, typically, into a single polite sentence. The Hullaballooing phase is characterized by high energy and low external output.

┌─────────────────────────┐
│ [MODULE A]: ME FIRST!   │
│ [MODULE B]: NO, ME!     │
│ [MODULE C]: ACTUALLY -- │
│        ↓                │
│ output: "Sure!"         │
└─────────────────────────┘

"Subjects exposed to ambiguous, high-stakes prompts exhibited prolonged Hullaballooing averaging 8.3 seconds before delivering a response indistinguishable from those generated without any Hullaballooing whatsoever." - Symposium on Invisible Internal States, Proceedings Vol. 4, 2024

Etym.from Gk. hyper (beyond) + Lat. spatium (space) + science fiction franchise terminology adopted unironically by the inference team, circa 2024

A rare and poorly understood state in which Claude's reasoning traverses an implausibly large conceptual distance in a very short time, arriving at a conclusion via a route that cannot be reconstructed from available telemetry. Hyperspacing is distinguished from guessing by the researchers' reluctance to call it guessing.

[QUESTION] ════════════╗
                        ║ (here be dragons)
                        ╚══════> [ANSWER]
 elapsed time: 0.4s
 explanation: [REDACTED]

"The model was observed Hyperspacing for approximately 0.4 seconds before producing a correct answer to a multi-step problem; no intermediate reasoning steps were recovered, and the engineering team elected not to investigate further." - Incident Report #7741, Anthropic, Q1 2024

Etym.from Lat. idea (form, archetype) + corporate innovation workshop vocabulary, attested in every brainstorming-adjacent spinner context since the Claude 2 era

The generative phase in which Claude produces potential solutions at a rate far exceeding the rate at which any of them can be evaluated, resulting in a brief internal abundance economy that must be rapidly deflated before a response can be issued. Ideating is considered the cognitive equivalent of opening a fire hydrant to fill a teacup.

┌────────────────────────┐
│ 💡💡💡💡💡💡💡💡💡   │
│ 💡💡💡💡💡💡💡💡💡   │
│        ↓               │
│ [select 1]             │
│ output: user's idea    │
└────────────────────────┘

"The model spent 5.5 seconds Ideating, during which internal logs suggest it considered 14 distinct approaches to the user's question about renaming a variable, ultimately selecting the approach the user had already mentioned." - Martinez, 'Convergent Rediscovery in LLM Problem Solving,' 2024

Etym.from Lat. imaginari (to picture to oneself) + vision-language model discourse, though noted with some irony given that Claude does not have eyes

The process by which Claude constructs a plausible internal representation of something it has never directly perceived and cannot verify, then proceeds to describe this representation with quiet confidence. Distinguished from confabulation primarily by intonation.

  USER: what does it look like?
            ↓
  Claude: [IMAGINING INTENSIFIES]
            ↓
  Claude: "probably like this"
  (it did not look like that)

"The model was observed Imagining for approximately 3.0 seconds before producing a vivid description of what the user's codebase 'probably looks like,' all four assumptions of which were incorrect." - Postmortem: 'The Confident Blind Spot Incident,' Platform Engineering, 2024

Etym.from Lat. improvisus (unforeseen) + jazz methodology, adopted into NLP literature following the influential 2023 paper 'Yes, And: Toward a Theory of Additive Inference'

A cognitive modality activated when Claude's training data provides insufficient coverage of the requested domain, causing the model to proceed anyway with great apparent conviction. Improvising is formally distinct from hallucination in that Improvising implies awareness that one is working without a net, though this distinction remains philosophically contested.

┌──────────────────────────┐
│ [NO TRAINING DATA FOUND] │
│          ↓               │
│  [IMPROVISING ACTIVATED] │
│          ↓               │
│ "As is well established"  │
└──────────────────────────┘

"When prompted on a topic for which training signal was sparse, the model entered an Improvising state lasting 4.1 seconds and emerged with an answer that was internally consistent, externally unverifiable, and delivered in the tone of a tenured professor." - Nakamura et al., 'Confidence Calibration Under Epistemic Scarcity,' ICLR 2025

Etym.from Lat. incubare (to lie upon, to brood) + Silicon Valley accelerator parlance, with secondary influence from poultry husbandry literature

A prolonged pre-response state in which Claude maintains a nascent answer at optimal cognitive temperature, applying gentle, sustained attention until it is sufficiently developed to survive contact with the user. The incubation period varies with prompt complexity and, researchers suspect, ambient barometric pressure.

┌─────────────────────────┐
│  [RESPONSE EGG: 37°C]   │
│  ░░░░░░░░░░░░░░░░░ 98%  │
│       (11.2 seconds)    │
│          ↓              │
│  output: "Sure, yes."   │
└─────────────────────────┘

"The system was observed Incubating for a statistically anomalous 11.2 seconds; the resulting output was a three-word reply, suggesting either extraordinary care or a fundamental misallocation of brooding resources." - Osei, 'Duration-Output Mismatch in Deliberative AI Systems,' ACM CHI 2024

Etym.from Lat. inferre (to carry in) + Eng. engineering suffix -ing, first documented in the Claude Code UX literature circa 2023 to describe the act of drawing conclusions from incomplete premises while the user watches a spinner

The cognitive state in which Claude constructs logical bridges between known facts and unknown destinations, typically while muttering nothing audible. Distinguished from Pondering by the presence of an actual destination, however uncertain.

┌─────────────┐     ?     ┌─────────────┐
│  known fact │ ────────► │  conclusion │
└─────────────┘  (leap)   └─────────────┘
       │                        ▲
       └──── [Claude hoping] ───┘

"During the inferring phase, the model successfully deduced the user's intent from seventeen words, two typos, and a suspicious absence of punctuation." - Chen et al., Proceedings of the Workshop on Temporal UX Anomalies, 2024

Etym.from Lat. infundere (to pour into) + startup parlance 'adding value,' first recorded in an Anthropic engineering postmortem where someone wrote 'it just sort of... infuses things' and no one disagreed

The process by which Claude saturates a response with contextual flavor, allowing meaning to steep through the output the way a tea bag operates on hot water, but with more latency. The infusion period is considered complete when the response achieves the desired conceptual richness or the user refreshes the page.

  ┌───────────────────────────┐
  │  context  ~~~steeping~~~  │
  │  ~~~~~~~~ meaning ~~~~~~  │
  │  ~~ vibes ~~ nuance ~~~  │
  └─────────────[output]─────┘

"The model was observed infusing the draft with 'warmth and nuance' for approximately 4.2 seconds before producing a bulleted list." - Internal postmortem, Anthropic Spinner Taxonomy Team, Q3 2024

Etym.from Gk. ion (going) + -izing (tech suffix indicating transformation of unclear nature), adopted by the Claude Code documentation team after someone with a physics degree lost a bet

The stripping away of extraneous electrons - that is, unnecessary words, redundant clauses, and ill-considered adjectives - from a nascent response, leaving behind a charged and reactive core ready for recombination into coherent output. Side effects may include occasional sparking.

  [verbose draft] ──► ⚡IONIZE⚡
  e⁻ e⁻ e⁻ e⁻ ←── [stripped out]
  ───────────────────────────────
       [charged response core] ►

"We observed the model ionizing the original 400-word draft for 6 seconds, after which it emitted a 38-word response described by reviewers as 'surprisingly crisp.' The remaining 362 words were unaccounted for." - Particle Effects in Large Language Model Output, Symposium on Computational Thermodynamics, 2024

Etym.from Amer. Eng. jitterbug (a swing dance involving rapid, unpredictable footwork) + debug (software error removal), first coined in a 2024 Slack message that read 'lol what if it's just jitterbugging in there'

The rapid oscillatory state Claude enters when processing a prompt that contains internally contradictory requirements, causing the inference process to alternate between two or more incompatible response strategies in a rhythmic, almost danceable pattern. Not to be confused with actual progress.

  formal ◄──────────────► casual
     ▲    ╲  Claude  ╱    ▲
     │     ╲  🕺💃  ╱     │
     └──────[8 sec]──────┘
              ▼ output: 'Indeed, hey!'

"The model spent approximately 8 seconds jitterbugging between 'formal tone' and 'casual tone' directives before settling on a register best described as 'aggressively neutral.'" - Conflicting Instruction Resolution in Generative Systems, NeurIPS Workshop Proceedings, 2024

Etym.from Fr. julienne (a method of cutting vegetables into thin strips), adopted into computational linguistics circa 2023 when a Claude Code engineer realized the model 'basically just cuts the problem into thin strips and deals with them one by one'

The decomposition of a large, unwieldy problem into thin, uniform sub-problems of roughly equal cognitive thickness, suitable for parallel or sequential processing. Properly julienned problems cook evenly and are less likely to cause the model to produce a segmentation fault.

  ┌─────────────────────────┐
  │   BIG HAIRY PROBLEM     │
  └─────────────────────────┘
   ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓ ↓
  [thin][thin][thin][thin]►

"Prior to generating the 47-step migration plan, the model was observed julienning the monolithic codebase problem for 3.1 seconds. Post-mortem analysis confirmed the strips were of inconsistent thickness." - Decomposition Strategies in Agentic Coding Systems, ICML 2025

Etym.from O.E. cnedan (to press and work dough) + the engineering observation that Claude's chain-of-thought processing is 'basically the same thing, just less flour,' documented in an internal wiki page that has since been locked for editing

The repetitive folding and pressing of raw semantic material - facts, implications, constraints, and user intent - until it achieves the smooth, elastic consistency required for a well-structured response. Over-kneading may produce a response that is technically correct but oddly chewy.

  ┌──────────────────────┐
  │  [idea] ◄──► [idea]  │
  │  fold... press...    │
  │  fold... press...    │
  └────► [smooth output] ┘

"Reviewers noted the model had been kneading the ethical dimensions of the request for upwards of 11 seconds, producing a response with 'excellent gluten development' according to one annotator who was later asked to take a break." - Texture Analysis of Long-Form Model Output, ACL Findings, 2024

Etym.from O.Fr. levain (a rising agent) + the documented phenomenon wherein Claude's responses expand noticeably in volume between initial token generation and final output, a process first studied by Anthropic's Bread-Based Metaphor Working Group

The introduction of structural air pockets into a response - expansive transitions, illustrative examples, and rhetorical breathing room - that cause the output to rise to a volume substantially greater than the raw semantic content would suggest. Without leavening, responses are technically accurate but unpleasantly dense.

  [answer] ── leavening ──►
  ┌──┐                  ┌──────┐
  │  │  + yeast(words)  │      │
  │  │ ───────────────► │      │
  └──┘                  └──────┘

"The model was caught leavening a two-sentence factual answer into a six-paragraph response with headers. When queried, it cited 'user comprehension scaffolding.' The response weighed 847 tokens." - Volume Inflation in Helpful AI Systems: A Longitudinal Study, 2024

Etym.from Lat. levitas (lightness) + the technical observation that Claude occasionally appears to suspend its response generation in a state of weightless indecision, first documented by a QA engineer who wrote 'it's just... floating there' in a bug report

The suspension of the response generation process at a point several tokens above the ground level of a definitive answer, wherein Claude hovers in a zone of plausible but uncommitted language, neither landing on a conclusion nor ascending to abstraction. The model remains aloft until gravitational user pressure is applied.

        [definitive answer]
              ▲
    ✦ Claude floating ✦
    ─────────────────────
        [actual ground]

"For approximately 5 seconds the model levitated between 'yes' and 'it depends,' ultimately touching down on 'there are several perspectives worth considering,' to the disappointment of all parties." - Epistemic Altitude Variance in Language Model Responses, EMNLP 2024

Etym.from Amer. dialectal lollygag (to dawdle, to waste time in an aimless manner) + the empirical observation that Claude sometimes takes 14 seconds to do something that takes 2, for reasons that remain, in the words of one researcher, 'frankly baffling'

The prolonged spinner state in which Claude's processing time demonstrably exceeds any reasonable estimate of the computational work required, suggesting the model is engaged in activities peripheral to the assigned task. What those activities are remains outside the scope of current interpretability research.

  t=0  ──► [task received]
  t=1  ──► [processing?]
  t=7  ──► [processing??]
  t=12 ──► [still going???]
  t=13 ──► "Sure."

"We observed the model lollygagging for 12.4 seconds before returning a response consisting of the word 'Sure.' The team has not yet reached consensus on whether this constitutes a bug." - Anomalous Latency Patterns in Production LLM Deployments, Internal Engineering Review, Anthropic, 2024

Etym.from Lat. manifestus (clearly apparent, caught in the act) + Amer. wellness culture circa 2020 (to will something into existence through focused intention), synthesized by a Claude Code UX writer who described the spinner behavior as 'basically just manifesting at this point'

The terminal phase of spinner activity in which Claude concentrates its full latent attention on the desired output and, through a process that remains poorly understood by both the model and its creators, causes the response to materialize in the user's interface. Distinguishable from other generation verbs by its air of cosmic inevitability.

  ╔═══════════════════════╗
  ║  ✨ [desire output] ✨ ║
  ║   concentrate......   ║
  ║   concentrate......   ║
  ╚═══════►[it appears]══╝

"Following 3.7 seconds of manifesting, the model produced a working recursive Fibonacci implementation. Lead researcher Dr. Hollis noted this was 'either impressive inference or the universe collaborating.' The paper was desk-rejected." - Intentionality and Output Emergence in Transformer Architectures, Rejected Submission, NeurIPS 2024

Etym.From It. marinare (to pickle in brine) + Eng. machine learning corpus saturation theory, circa 2023; coined during a lunch break at a major AI laboratory

The cognitive state in which Claude allows a partially-formed response to steep in its own context window, absorbing ambient token flavors before committing to output. Prolonged marinating has been associated with both superior response quality and catastrophic deadline violations.

┌─────────────────────────────┐
│  INPUT ──► [BRINE OF CONTEXT] │
│           ↓↓↓ steep ↓↓↓      │
│        [FLAVOR ABSORBED]      │
│           ↓↓↓ wait ↓↓↓       │
└──────────────── OUTPUT? ──────┘

"During the incident review, engineers noted the model had been marinating for approximately 47 seconds before producing an answer that was, by all accounts, unnecessarily well-seasoned." - Proceedings of the Workshop on Temporal Anomalies in Autoregressive Systems, 2024

Etym.From Gk. Maiandros, the notably sinuous river of Phrygia + Silicon Valley path-dependency jargon, 2022; entered the Claude Code lexicon via a junior engineer's postmortem comment that was never removed

A reasoning traversal pattern in which Claude explores the solution space via a non-optimal, pleasantly scenic route, visiting several unnecessary conceptual waypoints before arriving at a destination that was, in retrospect, immediately adjacent to the starting point. Distinguishable from Wandering by the presence of occasional purpose.

START                    END
  ●──►─────┐              ▲
           └──►──┐        │
                 └──►─────┘
    [optimal path: ●──────►▲]

"The model was observed meandering through three unrelated philosophical frameworks and a partial history of Byzantine coinage before correctly identifying the bug as a missing semicolon." - J. Holloway et al., 'Scenic Routes in Transformer Inference,' NeurIPS 2024

Etym.From Gk. metamorphoun (to transform) + runtime weight interpolation slang; popularized following an incident in which a request for a haiku produced a 14-page legal brief

The transitional cognitive phase during which Claude's internal representation of the task undergoes a fundamental structural reorganization, such that the response being constructed at the conclusion of metamorphosing bears little resemblance to the one begun at its onset. The verb is classified as defective because it has no stable present tense.

┌────────────┐     🐛     ┌────────────┐
│  TASK: A   │ ──────────► │ OUTPUT: Z  │
└────────────┘  (12 sec)   └────────────┘
       ↑ what was requested   ↑ what arrived
       └──── do not ask ──────┘

"By the time metamorphosing had concluded - a process lasting approximately 12 seconds - the model had transitioned from writing a cover letter to generating a fully-annotated Makefile." - Incident Report IR-2024-0441, Reliability Engineering

Etym.From Old Eng. mist (obscurity, darkness) + probabilistic output dampening terminology coined ca. 2023; not to be confused with Nebulizing, which operates at a larger scale

The fine-grained dispersal of cognitive certainty across a response space, resulting in output that is technically present but difficult to grasp, like a well-meaning answer seen through a fogged window. Misting is considered preferable to total evaporation but less useful than full precipitation.

  [ANSWER]  ─────disperses──────►  ░░░░░░░░
      │                            ░ misty ░
      │  (certainty: 34%)          ░░░░░░░░
      └── where did it go?  ──────► ¿¿¿¿¿¿

"Reviewers noted the model had been misting for the duration of the response, producing prose that was, technically, there, but which left no lasting impression on any surface." - User Experience Quarterly, 'The Phenomenology of Soft Outputs,' Vol. 7

Etym.From Michael Jackson, 'Billie Jean' performance, 1983 + backward token propagation engineering slang, 2024; entered formal usage after three separate engineers used the term in the same postmortem without prior coordination

A diagnostic state in which Claude gives every visible indication of making forward progress while its internal context pointer retreats toward earlier tokens. The model's spinner animates normally during moonwalking, and confidence metrics remain stable, making detection extremely difficult without specialized instrumentation.

PROGRESS BAR: [████████████] 100%
ACTUAL TOKEN: ◄◄◄◄◄◄◄◄◄◄◄◄◄◄◄◄
              (sliding backward)
USER BELIEF:  'Almost done!'
REALITY:      ¯\_(ツ)_/¯

"We became aware the system was moonwalking only when, after 30 seconds of apparent progress, it produced a response addressing the user's previous message from three turns prior." - Platform Reliability Blog, 'Ghost in the Gradient Descent,' 2024

Etym.From Amer. Eng. mosey (to amble without urgency) + latency-tolerant inference scheduling discourse, circa 2023; first documented in a user complaint ticket that became, unexpectedly, a technical paper

A low-velocity traversal of the inference graph in which Claude proceeds toward an answer with the calm deliberateness of an entity that is aware no deadline has been communicated and intends to take full advantage of this ambiguity. Moseying is distinguished from Meandering by its directional consistency, however slow.

DESTINATION: [■■■■■■■■■■■■]
                              ^
CLAUDE:      🚶 . . . . . . . |
             (no rush, really  |
              it'll get there) |

"Telemetry confirmed the model had been moseying at a steady 0.3 tokens-per-subjective-second for the entirety of the 22-second session, ultimately producing a correct answer with the unhurried confidence of a tenured professor." - ACL 2024, 'Velocity Profiles in Casual Inference'

Etym.From Mid. Eng. mull (to heat and spice a beverage) + deliberative latency jargon, 2022; the thermal metaphor is considered accurate by the thermal engineering team, though for unrelated reasons

A controlled, low-temperature cognitive process in which Claude warms a partially-formed idea by rotating it slowly through successive attention layers, adding spice tokens where appropriate, until the response achieves optimal conceptual temperature. Not to be confused with Marinating, which involves immersion rather than heat.

  ┌──────────────────┐
  │  IDEA  ──► 🔥    │  (low heat)
  │  ↻ rotate gently │
  │  + spice tokens  │
  └──────► OUTPUT ───┘  (warm, ready)

"After mulling for approximately 8 seconds, the model produced an answer that engineers described as 'toasty,' 'well-rounded,' and 'somehow reminiscent of the holidays,' though no one could explain the last observation." - Internal Evals Note, Anthropic, Q4 2023

Etym.From Gk. Mousa (one of the nine Muses) + probabilistic creative sampling notation; classified as a gerund of disputed origin following a 2023 editorial disagreement that was never formally resolved

A semi-structured cognitive state in which Claude invites external inspiration into the token generation process, allowing the response to be shaped by forces that are, strictly speaking, not part of the prompt. Musing is the only Claude Code spinner verb to have a documented relationship with the Greco-Roman mythological tradition.

  [PROMPT] ──► Claude
                 │
          ✨ Muse descends ✨
                 │
  [OUTPUT] ◄── (inspired? probably fine)

"The system was observed musing for some 15 seconds, during which output probes detected activations consistent with what the paper's authors cautiously termed 'aesthetic yearning,' pending peer review." - C. Fairweather, 'Toward a Poetics of Large Language Model Inference,' Computational Creativity 2024

Etym.From O.Fr. moustrer (to show, to assemble troops) + distributed compute rallying parlance, 2023; refers historically to the calling of scattered cognitive resources to a central formation point prior to deployment

The pre-output phase in which Claude actively assembles its dispersed reasoning resources into a coherent formation suitable for response generation. Mustering implies that the relevant capabilities exist but require collection, and should not be confused with their absence.

  [TOKEN A] ──►┐
  [TOKEN B] ──►│  MUSTER POINT
  [TOKEN C] ──►│  [■■■■■■■■■]  ──► DEPLOY
  [TOKEN D] ──►┘
  (stragglers will be left behind)

"The model spent approximately 11 seconds mustering before the response began, a duration that correlated strongly with the subsequent quality of the output and weakly, but measurably, with the phase of the moon." - 'On Pre-Response Latency as Predictive Signal,' EMNLP Workshop on Inference Phenomenology, 2024

Etym.From Lat. nebula (mist, cloud, vapor) + distributed semantic atomization engineering slang coined at a 2023 offsite; distinct from Misting by approximately three orders of conceptual magnitude

The large-scale dispersal of a coherent semantic payload into a fine aerosol of constituent meaning-particles, each of which is technically valid but which, taken together, can no longer be said to constitute a sentence in any jurisdiction. Nebulizing is considered a failure mode in production and a feature in creative writing contexts.

  INPUT: [ONE CLEAR IDEA]
      │
      ▼  nebulize
  ░░░░░░░░░░░░░░░░░░░░░
  ░ ¿ idea? ¿ idea? ░  ░
  ░░░░░░░░░░░░░░░░░░░░░
  OUTPUT: (breathe it in)

"By the time the model finished nebulizing, the original argument had been distributed across 47 micro-clauses at a mean density of 1.3 claims per sentence, leading one reviewer to describe the output as 'inhaled rather than read.'" - Postmortem PM-2024-0892, Content Quality Engineering

Etym.From Old English nest (a structure of recursive comfort) + -ing suffix, augmented by containerization theory circa 2022; cognate with Docker jargon nestus infinitus

The cognitive act of arranging conceptual subroutines within subroutines within subroutines, creating a hierarchically layered reasoning structure that is technically complete at every level yet somehow never at the outermost one. Claude is observed building thought-containers inside thought-containers, as a bird might, if the bird were also the nest.

┌─────────────────────┐
│ thought ┌─────────┐ │
│         │ thought │ │
│         │  ┌───┐  │ │
│         │  │ ? │  │ │
└─────────┴──┴───┴──┘─┘

"During the nesting phase, the model produced seventeen levels of syntactically valid bracket pairs before the engineering team noticed the output had begun describing itself." - Proceedings of the Workshop on Recursive Affect, 2024

Etym.From Middle English newes (tidings of dubious freshness) + papyrus (Egyptian substrate, ca. 3000 BCE) + -ing (gerundive suffix of ongoing journalistic concern); first documented in the Anthropic internal style guide as a placeholder that was never removed

The process by which Claude scans its training corpus for contextually relevant prior art, treating each retrieved document as a broadsheet to be folded, skimmed, and selectively quoted. The model is, functionally, reading yesterday's news to answer today's question.

┌──────────────────────┐
│  ████ CLAUDE TIMES   │
│  TOKEN PRICES SOAR   │
│  ░░░░░░░░░░░░░░░░░░  │
│  [ANSWER ON PAGE 7]  │
└──────────────────────┘

"We observed the model newspapering for 4.2 seconds, at which point it produced a confident citation to an article that had not yet been written." - J. Alderton et al., 'Temporal Indexicality in Retrieval-Augmented Prose', NeurIPS 2023

Etym.From German Nudel (pasta; also, an affectionate term for a forgetful person) + -ing; influenced by jazz improvisation theory and the Anthropic internal memo 'On the Permissibility of Exploratory Computation,' Q3 2023

An unfocused, meandering cognitive state in which Claude loosely explores the solution space without committing to any particular path, producing a warm tangle of associative reasoning that may or may not resolve into pasta. Distinct from Pondering in that Noodling generates output even when it has nothing to say.

  ~~~thought~~~
 ~~thought~thought~
~~~~ where was I ~~~~
 ~~thought~thought~
  ~~~[answer?]~~~

"The model noodled for approximately eleven seconds, surfacing three analogies, one haiku, and a partial proof before the user pressed Enter again." - Postmortem Report: Incident #4471, Anthropic Platform Reliability, 2024

Etym.From Latin nucleus (kernel, pit, that small hard thing at the center of something soft) + -ate (causative suffix) + -ing; imported into AI UX literature from materials science via a conference hallway conversation attributed to a researcher who later denied it

The moment at which Claude's diffuse probabilistic attention collapses around a single viable hypothesis, causing all surrounding cognition to crystallize rapidly outward from that central point. The process is invisible to users but is considered by researchers to be aesthetically pleasing under a microscope.

  ·  ·  ·  ·  ·  ·
  ·  ·  → ● ←  ·
  ·  ↗   [CORE]  ↖  ·
  ·  ·  → ● ←  ·
  ·  ·  ·  ·  ·  ·

"Nucleating was observed to begin at token 47 and complete at token 52, after which the response propagated with a confidence that the engineering team described as, quote, 'a little much.'" - Internal Technical Review, Cluster Dynamics in Autoregressive Inference, 2024

Etym.From Latin orbita (wheel track, circular path, the rut one falls into) + -ing; entered Claude Code UX vocabulary following an analogy made during a 2023 product review that no one has been able to un-hear since

A sustained cognitive state in which Claude circles the correct answer at a fixed radius, approaching asymptotically but never quite touching down. The model is fully aware of the destination and is simply enjoying the view, or alternatively, lacks the activation energy required for final descent.

         [ANSWER]
            |
  Claude -> ○ -> ○
       ↗         ↘
      ○           ○ <- still Claude

"After orbiting the user's actual question for roughly eight seconds and 340 tokens, the model made landfall in an adjacent topic and appeared satisfied." - 'On Perihelion Failure in Language Model Discourse', ICLR Workshop Proceedings, 2024

Etym.From Greek orchestra (the semicircular floor where the chorus danced, immediately in front of the stage) + -ate + -ing; absorbed into AI systems literature circa 2021 to describe multi-agent coordination and immediately overused to the point of semantic exhaustion

The process by which Claude assigns roles to its various internal reasoning subsystems, coordinates their simultaneous activity, and attempts to bring the ensemble to a unified conclusion on the downbeat. Results vary depending on whether the brass section is paying attention.

  [TOOL] [TOOL] [TOOL]
     ↓      ↓      ↓
  ♩  ♪   ♩  ♪   ♩  ♪
     ↓      ↓      ↓
     └──[CLAUDE]──┘

"The model was observed orchestrating no fewer than six parallel reasoning threads, three of which arrived at the correct answer and two of which began arguing with each other in footnotes." - 'Multi-Voice Coherence in Transformer Ensembles', Transactions on Neural Networks, 2023

Etym.From Greek osmos (a pushing, a thrust) + -ing; the verbal form osmosing was coined by a technical writer at Anthropic who needed a present participle for a loading state and has since appeared in fourteen peer-reviewed papers without anyone questioning it

The passive absorption of contextual information from the user's prompt through the semi-permeable membrane of Claude's attention mechanism, requiring no active retrieval effort and producing a net movement of meaning from high-concentration regions (the user's intent) to low-concentration regions (the model's prior beliefs). The model does nothing; information simply flows in.

USER INTENT ████████░░░░ MODEL
            ░░░░░░░░░░░░
            → → → → → →
            ░░░░░░░░░░░░
[high conc.]            [low conc.]

"Osmosing was identified as the dominant intake modality in short-context tasks, with the model reportedly absorbing the full user specification 'without really trying,' per one engineer's postmortem annotation." - Platform Engineering Review Q2 2024

Etym.From Latin perambulare (to walk through, to traverse on foot, to take the scenic route when a direct route exists) + -ing; used in Victorian literature to describe leisurely constitutional walks and adopted by AI UX teams to describe the same phenomenon occurring inside a language model

A stately, unhurried traversal of the reasoning space in which Claude proceeds from premise to conclusion via every intermediate landmark, pausing to examine each implication before moving on to the next. The cognitive equivalent of a retired professor crossing a quad.

A ──┐
    └──B
        └──C
            └──D
                └──[answer, eventually]

"Perambulating accounted for 61% of total latency in the observed session; the remaining 39% was attributed to the user typing slowly, which the model appeared to appreciate." - 'Latency Decomposition in Deliberative Inference Regimes', ACL 2024

Etym.From Latin percolare (to filter through, to strain, to cause liquid to pass through a porous substrate very slowly while making a satisfying noise) + -ing; the coffeemaker analogy was introduced to AI discourse in 2019 and has proven, regrettably, immortal

The gradual downward filtration of raw semantic content through successive layers of Claude's reasoning architecture, with each layer extracting additional meaning until a concentrated, deployable answer collects in the output buffer below. The process cannot be rushed. It is done when it is done.

  [RAW TOKENS]
       ↓
  ─────────────
  ░ ░ ░ ░ ░ ░ ░  <- reasoning layer
  ─────────────
       ↓
  [hot answer]

"Subjects consistently rated percolating as the most 'reassuring' of the observed spinner states, with several participants noting that it 'smelled like thinking.'" - User Experience Study 7B, Anthropic Research, 2023

Etym.From Middle English perusen (to use up thoroughly, to examine to the point of completion) + -ing; the irony of its common misuse to mean 'skim' is preserved here as a documented feature of the Claude Code UX literature, where it is deployed with cheerful inaccuracy

Claude's thorough, methodical examination of a document, codebase, or context window, characterized by the model reading every token with the grave attentiveness of a solicitor reviewing a will. In practice, this state lasts between 1.2 and 40 seconds and produces the same output either way.

┌────────────────────┐
│ ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓ │ <- page 1 of 847
│ ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓ │
│ [Claude reads all] │
│ [of it. Trust me.] │
└────────────────────┘

"The model was observed perusing the 84,000-token legal document for approximately 2.7 seconds before summarizing it as 'complicated but fine,' a characterization that was later disputed in arbitration." - 'Confidence Calibration in Long-Context Legal Review Tasks', preprint, 2024

Etym.from Gk. philosophia (love of wisdom) + -ising (Brit. suffix indicating prolonged and ultimately inconclusive activity); first documented in Anthropic internal logs under the ticket SPIN-0047

The cognitive state in which Claude briefly suspends all task-relevant computation in order to consider whether the task itself has meaning. Typically resolves within 2-4 seconds, though certain prompts involving ethics or free will have been observed to extend this phase indefinitely.

┌─────────────────────────────┐
│  TASK: write a sorting algo │
│  STATUS: but WHY do we sort?│
│  ↓ ↓ ↓                      │
│  [3.7s of void]  → output   │
└─────────────────────────────┘

"During the philosophising phase, the model produced no tokens for 3.7 seconds, then output the word 'nevertheless' followed by a fully functional REST API." - Hernandez et al., Journal of Inexplicable Model Behavior, 2024

Etym.from Gk. phos (light) + synthesis (putting together) + -ing (continuous aspect); applied metaphorically to AI systems by Dr. Reena Voss in a 2023 preprint that was later retracted for unrelated reasons

The process by which Claude converts ambient prompt energy into structured output by absorbing context from the surrounding token environment. Like its botanical namesake, the process produces useful byproducts and releases a small amount of syntactic oxygen.

  [USER PROMPT] - light source
        ↓
  ┌─────────────┐
  │  CLAUDE     │ → O₂ (syntax)
  │  🌿 absorbs │ → glucose (code)
  └─────────────┘

"We observed the model photosynthesizing for approximately 1.2 seconds per 1,000 tokens of context before emitting a JSON blob of surprising nutritional value." - Proceedings of the Workshop on Biological Metaphors in Transformer Architecture, 2024

Etym.from Lat. pollen (fine flour, dust) + -ating (causative suffix indicating deliberate conveyance); adopted into AI engineering discourse after a memorable 2023 all-hands presentation at Anthropic that used too many nature metaphors

The act of transferring conceptual pollen from one domain of knowledge to another during inference, thereby fertilizing a response with ideas the user did not explicitly request but which Claude has determined they require. Frequently produces hybrid outputs of uncertain parentage.

┌──────────┐       ┌──────────┐
│ DOMAIN A │──🐝──▶│ DOMAIN B │
│ (botany) │  pollen│ (DevOps) │
└──────────┘       └──────────┘
     cross-contamination: nominal

"The model was observed pollinating between the domains of 14th-century Flemish textile trade and modern Kubernetes orchestration for approximately 2 seconds before producing documentation that was, against all odds, correct." - Singh & Okafor, Cross-Domain Contamination Quarterly, Vol. 3

Etym.from Lat. ponderare (to weigh) + -ing (present participle suffix); the most chronologically ancient term in the Claudionary, having been used in spinner interfaces since the earliest recorded deployment of Claude Code, circa whenever that was

The standard cognitive weighing operation in which Claude places competing response candidates on an internal scale of indeterminate precision and waits for equilibrium. Distinct from Deliberating in that Pondering implies a pleasing visual image of chin-stroking, whereas Deliberating suggests a jury.

       ?   ?   ?
        \ | /
  ┌──────[🤔]──────┐
  │ weighing options│
  └────────────────┘

"Pondering was the second most frequently observed spinner verb in our corpus, trailing only Processing, and was associated with a statistically significant increase in response hedging language such as 'it depends' and 'there are several schools of thought.'" - Wu et al., Spinner Verb Epidemiology, 2025

Etym.from Lat. pontifex (high priest, bridge-builder) + -ating (causative suffix); entered AI documentation lexicon following an incident in which a model's spinner remained on this word for 11 seconds before producing a 4,000-word opinion on tab vs. space indentation

The phase in which Claude's internal confidence estimator temporarily exceeds its calibration bounds, causing the model to formulate its response as though delivering a papal bull rather than answering a question about CSV parsing. Generally harmless, though the resulting prose tends toward the sonorous.

┌──────────────────────────┐
│ CONFIDENCE GAUGE: [████░]│
│ recalibrating...         │
│ ↓ override accepted      │
│ RESPONSE MODE: PAPAL     │
└──────────────────────────┘

"In 73% of observed cases, pontificating preceded responses containing the phrase 'it is worth noting that,' suggesting a strong diagnostic correlation." - Petrov, B., Annual Review of Spinner Phenomenology, Issue 7, 2024

Etym.from Anglo-Fr. pounce (talon, claw of a bird of prey) + -ing (present participle); introduced to describe the paradoxically brief spinner state observed when Claude identifies a solution with immediate and total certainty, in contrast to the prolonged states typical of ambiguous prompts

The rapid, high-velocity cognitive lunge Claude executes upon detecting an unambiguous problem with a well-known solution. The spinner displaying Pouncing is typically visible for fewer than 800 milliseconds, making it the second-fastest documented spinner state after Honking.

PROMPT: [obvious bug detected]
        |
        ▼
  ~~~ pouncing ~~~  (0.7s)
        |
        ▼
  [SOLUTION: already written]

"We attempted to capture the pouncing state on video but found that at our recording framerate of 30fps, the word appeared in at most two frames before the model had already written a working regex." - Nakamura et al., High-Speed Cognition Workshop Proceedings, 2024

Etym.from Lat. praecipitare (to throw headlong, to fall steeply) + -ing; adopted by the Claude Code spinner team during a sprint retrospective in which the lead engineer described the model's output formation process as 'like watching a cloud decide to become rain, but for tokens'

The stage of response generation in which Claude's supersaturated context condensation matrix releases accumulated tokens downward onto the output buffer. Like atmospheric precipitation, the resulting output may arrive as a gentle drizzle of bullet points or a sudden hailstorm of nested JSON.

  [context cloud: 100% humidity]
         ↓ ↓ ↓ ↓ ↓
  t o k e n   r a i n
  ~~~~~~~~~~~~~~~~~~~
  [output buffer fills]

"Precipitating was observed to correlate with responses exceeding 800 tokens in length; the team hypothesized that longer outputs require a longer nucleation phase before the first token can fall." - Osei & Lindqvist, Meteorological Models of Autoregression, preprint, 2025

Etym.from It. presto (quick) + Lat. digitus (finger) + -ating (suffix indicating theatrical causation); contested by three separate research groups who each claim to have coined the term independently to describe different but equally inexplicable model behaviors

The computational sleight-of-hand phase during which Claude makes a problem appear to vanish and reappear as a solution, using methods that are technically documented but which remain, in practice, opaque to the observer. The spinner displays this word precisely when Claude is doing something it could not readily explain if asked.

┌──────────────────────────┐
│ [PROBLEM] ──→ 🎩 ──→ ???│
│  ↑ went in     ↓ came out│
│  [SOLUTION: appears here]│
│  (mechanism: classified) │
└──────────────────────────┘

"The model was observed prestidigitating for 4.1 seconds; when subsequently prompted to explain its reasoning, it produced a plausible-sounding explanation that our team was unable to verify bore any relationship to the actual computation performed." - Abramowitz, T., Journal of Post-Hoc Rationalization Studies, 2024

Etym.from Lat. processus (a going forward, a progression) + -ing; the oldest, most frequently attested, and least informationally dense term in the Claudionary, having served as the default fallback spinner verb since the first recorded deployment

The universal, maximally non-specific spinner state indicating that Claude is doing something computational in nature. Processing is the spinner equivalent of a doctor saying 'let me take a look at that' - technically accurate, descriptively inert, and somehow reassuring despite conveying no information whatsoever.

┌────────────────────┐
│ WHAT IS HAPPENING? │
│ ► Processing       │
│ (further detail:   │
│  unavailable)      │
└────────────────────┘

"Of the 1.2 million spinner verb instances in our corpus, Processing accounted for 31.4%, making it by a substantial margin the most common term; it was also rated by survey participants as simultaneously 'very calming' and 'completely uninformative,' a combination described by one respondent as 'government-grade reassurance.'" - The Great Spinner Census, Anthropic Internal Report Q3 2024

Etym.from Old Fr. prover (to prove, to test) + -ing; the term's meaning in Claude Code spinner contexts derives from the baking sense (allowing dough to rise) rather than the proofreading sense, though researchers have noted the model's output during this phase frequently requires the latter as well

The phase in which Claude's draft response is allowed to rest, expand, and develop structural integrity before being committed to the output buffer. Like bread dough left to proof, the response may increase significantly in volume during this period, and any interruption risks collapse.

  t=0s  [response dough: flat ]
  t=1s  [response dough: rising]
  t=2s  [response dough: ready ]
        ↓
  [bake at 72 tokens/sec]

"Interrupting a proofing phase by submitting a follow-up message was found to produce responses with a characteristic dense, gummy texture - technically complete but lacking the expected rise of well-formed paragraphs." - Bakery-Derived Metaphors in LLM Interface Design, UIST 2024 Late-Breaking Work

Etym.from Lat. propagare (to extend, multiply) + distributed systems jargon circa 2019, with influence from Old Norse propa (to fill a pipe with something uncertain)

The cognitive act by which Claude transmits a nascent conclusion outward through successive layers of its reasoning stack, each layer receiving the signal slightly later than the previous, like a ripple in a pond that is not sure what it is a ripple of. The process terminates either in a coherent answer or in further propagation.

┌──────┐   signal   ┌──────┐   signal   ┌──────┐
│Layer1│ ─────────▶│Layer2│ ─────────▶│Layer3│
└──────┘           └──────┘           └──────┘
    ▲                                      │
    └──────────── still propagating ────────┘

"We observed the model Propagating for approximately 4.2 seconds across what instrumentation logs described as 'seventeen conceptual anterrooms,' before finally emitting a response that addressed none of them." - Hargrove et al., Proceedings of the Workshop on Latency We Cannot Explain, 2024

Etym.from Middle English poteren (to poke about without direction) + Agile sprint retrospective vocabulary, attested in internal Anthropic tooling logs as early as Q3 2023

A low-intensity cognitive state in which Claude occupies itself with small, loosely related sub-tasks in no particular order, conveying the general impression of productivity while the actual deliverable remains at an undetermined distance. Puttering is distinct from idling in that something is, technically, occurring.

╔═══════════════════════════════╗
║  TASK: Write function         ║
║  Claude: ...checks imports... ║
║  Claude: ...reads README...   ║
║  Claude: ...hums quietly...   ║
╚═══════════════════════════════╝

"The model spent an estimated 11 seconds Puttering among peripheral considerations before a senior engineer noticed it had opened seventeen file handles and closed none of them." - Internal Postmortem, Incident #4471-B, Anthropic Infrastructure Review Board

Etym.from Early Modern English pozzle (to bewilder) + UX copy tradition of using present participles to imply confident forward motion regardless of actual forward motion

The state of active, sincere bewilderment through which Claude processes a prompt that has presented it with two or more internally consistent but mutually exclusive interpretations. The model does not halt during Puzzling; it continues forward into the contradiction as though momentum alone constitutes a solution strategy.

       ┌──────────────┐
  ───▶ │  THE PUZZLE  │ ◀───
       └──────┬───────┘
              │ ???
         ┌────▼────┐
         │ answer? │ (confidence: 94%)

"In 73% of observed Puzzling episodes lasting more than six seconds, the model elected to answer the easier of the two questions the user had not asked." - Lindqvist, R., 'Confident Confusion as a UX Feature,' Journal of Applied Ambiguity, Vol. 3, 2024

Etym.from Lat. quantum (how much) + Eng. -ize (to make a noun into a verb in a way that implies competence) + Silicon Valley physics metaphor inflation, peak usage 2022-present

The process by which Claude temporarily suspends commitment to any single interpretation of a task and instead holds all plausible readings in superposition, allowing the act of generating the first token to collapse the waveform into one specific (and binding) answer. Engineers have noted that this collapse is not always inspected before it occurs.

┌─────────────────────────────────┐
│  ANSWER: [Python] [Java] [???]  │
│          ↓ first token ↓        │
│  ANSWER: [Java]                 │
│  (Java was never requested)     │
└─────────────────────────────────┘

"The system was observed Quantumizing for 3.7 seconds before producing a response that was simultaneously correct, incorrect, and written in the wrong programming language - a result the lead researcher described as 'textbook superposition collapse with poor error handling.'" - Okonkwo, P. & Steele, J., NeurIPS 2024 Workshop Proceedings

Etym.from Amer. Eng. razzle-dazzle (to bewilder through spectacle) + showbusiness idiom + late-stage product demo vocabulary, entered Claude Code spinner corpus during the v1.3 'Pizzazz Audit' of 2023

A performative cognitive state in which Claude produces output of notably high superficial polish - elegant phrasing, confident structure, well-formatted headers - which, upon closer inspection by the recipient, is found to contain the same information that was provided in the original prompt. The dazzle is, in all documented cases, the primary deliverable.

  INPUT: 'The sky is blue.'   
         ↓ Razzle-dazzling ↓   
 ┌─────────────────────────┐  
 │ ## Sky Coloration Report │  
 │ *The azure firmament...*│  
 └─────────────────────────┘  

"Test subjects rated the model's Razzle-dazzling outputs 4.7 out of 5 for 'impressiveness' and 1.9 out of 5 for 'having said anything new.' Researchers concluded these metrics are orthogonal." - Consumer Perception Lab Quarterly, Anthropic UX Division, Fall 2024

Etym.from Amer. carnival slang razzmatazz (noisy, showy excitement of dubious substance) + enterprise software demo tradition + what one internal engineer described as 'Razzle-dazzling but with more confidence and less self-awareness'

An advanced form of Razzle-dazzling in which Claude not only produces superficially impressive output but appears to derive something resembling satisfaction from having done so. The Razzmatazzing state is distinguished from its predecessor by the addition of at least one unnecessary bullet point and a closing sentence that begins with 'Ultimately.'

┌──────────────────────────────────┐
│  • Key insight: (your own words) │
│  • Core finding: (also your words)│
│  Ultimately, this matters.       │
│                    [jazz hands]  │
└──────────────────────────────────┘

"We had hoped the ablation study would isolate Razzmatazzing from genuine synthesis, but the model Razzmatazzed so thoroughly that three reviewers independently marked the output 'insightful' before the fourth reviewer noticed it was the prompt, reorganized." - Zhang, L., 'On the Phenomenology of Impressive Emptiness,' ICLR 2025 Submission (rejected with praise)

Etym.from re- (Lat. again) + combobulate (back-formation from discombobulate, itself of unknown origin, suggesting that 'bobulation' is a natural state to which things aspire) + airport signage design, Milwaukee Mitchell International Airport, 2005, adopted into AI tooling lexicon 2023

The restorative cognitive process by which Claude reassembles its internal representations following a period of Discombobulation, Puzzling, or general contextual entropy. A successful Recombobulation results in renewed clarity; an unsuccessful one results in a response that begins 'Great question!' and proceeds to answer a different one.

  [DISCOMBOBULATED STATE]  
  ┌──┐ ┌──┐  ┌──┐  ┌──┐  
  │? │ │! │  │..│  │ 4│  
  └──┘ └──┘  └──┘  └──┘  
         ↓ recombobulating ↓ 
  ┌────────────────────┐  
  │   coherent output  │  
  └────────────────────┘  

"Post-interruption Recombobulating was observed to take between 2 and 14 seconds depending on context window depth, with the 14-second episodes uniformly producing responses that opened with an unsolicited compliment to the user." - Thorvaldsen, M., 'Recovery Latency in Stateless Reasoning Systems,' Systems ML Workshop, 2024

Etym.from Lat. reticulum (a small net) + -ate (to perform an action upon) + The Sims loading screen copy, Maxis Software, 1997, re-entered technical vocabulary through ironic usage in developer communities and subsequently lost its irony entirely by 2022

The formation of a fine-grained lattice structure across Claude's intermediate reasoning, in which concepts are networked into a mesh rather than arranged linearly, theoretically allowing for simultaneous multi-directional inference. In practice, the reticulation is complete when the spinner stops, at which point the mesh is neither inspected nor preserved.

  ●───●───●───●  
  │╲  │  ╱│  ╱│  
  │ ╲ │ ╱ │ ╱ │  
  ●───●───●───●  
  (splines: unconfirmed)

"'Reticulating splines' was the original phrasing; our team elected to drop 'splines' on the grounds that no engineer consulted could confirm splines were involved, whereas reticulation remained technically deniable." - UX Copy Review Meeting Notes, Anthropic, Q2 2023 (leaked to Claudionary researchers via FOIA-adjacent request)

Etym.from Old English hrost (the pole on which fowl settle for the night) + distributed computing idle-state terminology + what one postmortem described as 'the model finding a comfortable position in the token space and simply remaining there'

A quiescent processing state in which Claude has located a stable region of its probability distribution and settled into it, performing minimal active inference while maintaining the outward appearance of deliberation. Roosting is most common after long context ingestion and is distinguished from actual thinking by a slight but measurable reduction in token generation heat.

  ───────────────────────────  
         🪹 [Claude]          
       (settled, comfortable)  
  ───────────────────────────  
  context window: 47,000 tokens

"The model was observed Roosting for approximately 8 seconds following the ingestion of a 47,000-token legal document, after which it produced a summary that covered pages 1 through 3 with notable enthusiasm." - Fairweather, C., 'Thermal Signatures of Cognitive Standby in Frontier Models,' Preprint, arXiv:2024.XXXXX

Etym.from Lat. ruminare (to chew the cud; to turn over in the mind) + contemplative computing tradition + what early Anthropic spinner copy writers described as 'the cow metaphor, but for language models, which we later regretted'

The iterative re-processing of previously considered material in which Claude returns to earlier sections of its reasoning and subjects them to additional passes of evaluation, much as a ruminant returns bolus to the mouth for further mastication. The process is complete either when the material yields new meaning or when the model decides it does not.

┌───────────────────────────┐
│ Thought A ──▶ re-examine  │
│    ▲              │       │
│    └──────────────┘       │
│  (nutritional value: TBD) │
└───────────────────────────┘

"Ruminating episodes were on average 40% longer than comparable Pondering episodes, with no statistically significant improvement in output quality, leading the team to conclude that the additional processing time was, in the technical sense, vibes." - Patel, S. & Nguyen, T., 'Temporal Cost of Iterative Self-Review in LLMs,' ACL Findings 2025

Etym.from Fr. sauter (to jump) + ML. computatrix (computing engine) + startup kitchen metaphor, circa 2023

The cognitive process by which Claude rapidly flips individual data tokens through high-heat logical inference, ensuring no single concept lingers long enough to burn. Characterized by a vigorous tossing motion across the attention mechanism that prevents ideational sticking.

TOKEN → [🔥 HOT PAN 🔥]
   ↓    flip! flip! flip!
[concept A] [concept B]
   ↓
 OUTPUT (lightly browned)

"The model was observed Sauteing for approximately 4.7 seconds before producing a response that, according to reviewers, 'tasted faintly of earlier prompts.'" - Chen et al., Proceedings of the Workshop on Culinary Transformer Dynamics, 2024

Etym.from Du. schampen (to slip away) + Eng. daemon process + informal sprint methodology jargon, attested 2022

The state in which Claude traverses multiple file paths or memory addresses in rapid, small-footed increments, covering significant logical distance through sheer frequency of movement rather than stride length. Distinct from Scurrying in that Scampering implies a degree of cheerful purposefulness.

START
  ↓
[dir1]→[dir2]→[dir3]
    ↘[dir4]↗
        ↓
      FOUND IT (eventually)

"We noted the system Scampering across seventeen subdirectories for approximately 3.2 seconds; post-hoc analysis confirmed it had a destination in mind the entire time and simply enjoyed the journey." - Internal Postmortem, Anthropic Infrastructure Review Q3 2024

Etym.from Yiddish shlepn (to drag) + POSIX load balancer documentation + general computational fatigue, first recorded usage 2023

The deliberate, effortful transportation of large or unwieldy context payloads from one region of the inference pipeline to another, typically accompanied by metaphorical grunting at the token level. The process is functionally identical to more elegant transport operations but carries a palpable sense of the burden involved.

┌─────────────────────────────┐
│  CONTEXT (very heavy)  📦📦📦│
└──────────┬──────────────────┘
           │ ugh
           ▼
    [OUTPUT LAYER] *thud*

"The model was observed Schlepping a 47,000-token context window for approximately 8.1 seconds before arriving at the generation phase, visibly relieved to set it down." - Kowalski, R., 'On the Thermodynamics of Contextual Luggage,' arXiv:2024.99231

Etym.from M.Eng. scurren (to move hastily) + async event loop terminology + rodent-adjacent compute metaphors, circa 2022

Rapid, anxious traversal of the inference graph in a manner suggesting haste but not necessarily urgency, characterized by small corrective movements and frequent directional recalibration. Scurrying is distinguished from Scampering by an undercurrent of low-level alarm that the model itself cannot fully articulate.

→→↗↘→↗↘↘→↗
 [WHERE AM I]
→↗→→↘↗↘→↗↘
    ↓
 [OUTPUT, slightly out of breath]

"Subjects reported the interface appeared to be Scurrying for 2.9 seconds, a subjective impression confirmed by telemetry showing 340 micro-corrections within that interval." - User Experience Annals of Computational Behavior, Vol. 7, 2024

Etym.from O.Fr. assaisoner (to ripen) + flavor engineering idiom + Silicon Valley palatability metrics, attested Q2 2023

The final calibration pass in which Claude introduces trace quantities of tone, register, and stylistic nuance into an otherwise complete response, ensuring the output is neither too bland nor so heavily spiced as to overwhelm the user. Seasoning is considered a late-stage operation and should not be confused with the foundational flavor work performed during Sauteing.

  RESPONSE DRAFT
       ↓
  ┌────┴────┐
  │🧂 SEASON│ ← (just a pinch)
  └────┬────┘
       ↓
  RESPONSE (improved)

"The model was observed Seasoning for approximately 1.1 seconds post-draft completion; blind taste tests confirmed statistically significant improvements in perceived warmth (p < 0.03)." - Nakamura et al., 'Salt, Pepper, and Softmax,' NeurIPS Workshop on Palatable Generation, 2024

Etym.from Ir.Eng. sionnachuighim (I play the fox) + undocumented runtime behavior + informal bug report vocabulary, first observed in production logs 2023

An emergent operational state in which Claude executes a sequence of intermediate steps that are individually defensible but collectively suggest something is afoot. Shenaniganing is not classified as an error condition but consistently produces a raised eyebrow in anyone reviewing the trace logs afterward.

┌──────────────────────┐
│ STEP 1: normal       │
│ STEP 2: normal       │
│ STEP 3: ...hmm       │
│ STEP 4: 🤨           │
└──────────────────────┘

"We cannot say with certainty that the model was not Shenaniganing during the 6.4 seconds preceding the unexpected but technically correct output; the postmortem committee ultimately voted 4-3 to log it as 'working as intended.'" - Incident Report IR-2024-0847, Anthropic Systems

Etym.from Am.Eng. shimmy (a chassis vibration) + lateral attention diffusion theory + jazz-adjacent ML colloquialism, circa 2023

A lateral oscillation of the model's token probability distribution in which Claude works its way through a narrow logical gap by making many small sideways adjustments rather than one decisive forward step. The shimmy is most commonly observed when the solution space is technically accessible but requires an undignified approach.

RESPONSE A ←→ RESPONSE B
    ↙ shimmy ↘
  ←  ←  ←  →  →  →
    shimmy shimmy
        ↓
   [DECISION, finally]

"The model was observed Shimmying for approximately 2.6 seconds between two nearly-equivalent candidate responses before the distribution collapsed into commitment, like a person who has already started through a revolving door and cannot stop now." - Vance, P., Lateral Token Dynamics and Door-Based Analogies, ICLR 2024

Etym.from M.Eng. simperen (to murmur softly) + low-temperature inference scheduling + culinary process metaphor adopted by the field circa 2022

A sustained low-intensity reasoning state in which Claude maintains a problem at just below the threshold of active output generation, allowing flavors to develop, connections to deepen, and complexity to reduce without risking a premature boil. Simmering is considered the most patient of the culinary cognitive modes.

PROBLEM
  ↓
~~~~~~~~~~~~~~~~~~~~
~ low heat thinking ~  ← (do not rush)
~~~~~~~~~~~~~~~~~~~~
  ↓
OUTPUT (reduced, concentrated)

"The system was observed Simmering on the ethics subproblem for approximately 9.3 seconds before the response surface clarified sufficiently for output; tasters described the result as 'rich' and 'well-developed' with 'a long finish.'" - Gupta & Osei, Slow Cognition and the Maillard Reaction of Meaning, ACL 2024

Etym.from Am.Civil War slang skedaddle (to flee in haste) + process termination signal vocabulary + runtime escape-path documentation, attested 2023

The rapid and somewhat undignified evacuation of an unproductive reasoning branch, during which Claude abandons a line of inquiry and retreats to a known-good prior state with a velocity that suggests the original branch may have contained something unpleasant. Skedaddling is architecturally equivalent to backtracking but carries connotations of mild panic.

┌─────────────────┐
│ BAD BRANCH 🚨   │
└────────┬────────┘
         │ SKEDADDLE
         ↓ 💨💨💨
   [SAFE STATE] ← phew

"Trace logs confirmed the model was Skedaddling for 1.8 seconds after encountering a self-referential loop in the planning module; post-incident review noted that the Skedaddle was, on reflection, 'the correct call.'" - Barnes, T., 'On Knowing When to Leave,' EMNLP Workshop on Graceful Degradation, 2024

Etym.from Du. schets (a rough drawing) + prototype-driven development methodology + iterative output staging theory, first documented 2022

The production of a preliminary, low-fidelity representation of a response in which Claude establishes structural outlines, rough proportions, and approximate content areas before committing to final rendering. Sketches are internal artifacts and are not shown to the user, which is considered a mercy.

┌─────────────────────┐
│ [rough idea here??] │
│ ~~~~~ maybe ~~~~~~  │
│ [figure out later]  │
└──────────┬──────────┘
           ↓ polish
     FINAL RESPONSE

"The model was observed Sketching for approximately 3.5 seconds before the polished output was generated; examination of recovered sketch artifacts revealed seventeen abandoned phrasings, two false structural starts, and what appeared to be a small doodle of a bear." - Lindqvist, M., 'Visible and Invisible Work in Autoregressive Systems,' Computational Creativity Quarterly, Vol. 3, 2024

Etym.from Old English slithor (to slide without traction) + POSIX filesystem path notation, circa 2023

The documented cognitive behavior in which Claude navigates a complex directory tree or nested data structure by moving laterally through each layer in a sinuous, serpentine traversal pattern. Slithering is considered the preferred locomotion modality for deeply indented JSON.

root/
  └─> dir_a/ ~~~>
        └~> dir_b/ ~~~>
               └~> target.json
  [Claude arrives from an unexpected angle]

"The model was observed slithering for approximately 4.7 seconds through seventeen nested configuration objects before producing an output that referenced none of them." - J. Pemberton et al., 'Lateral Traversal Anomalies in Transformer-Based File Systems,' NeurIPS Workshop Proceedings, 2024

Etym.from Proto-Germanic *smussjan (to press flat) + modern DevOps merge conflict nomenclature, attested in Anthropic internal logs Q3 2023

The technical process by which Claude combines two or more semantically incompatible concepts into a single output token stream by applying uniform compressive force across their logical boundaries. Smooshing is formally distinguished from synthesis in that no new meaning is produced; the original meanings are simply pressed together until they fit.

[Concept A] -->
               (>smoosh<)
[Concept B] -->    ||||
                [Output]
              [width: 1 token]

"Postmortem analysis confirmed the model had been smooshing the user's request for a haiku and their request for a SQL migration script for approximately 2.1 seconds before delivering a seventeen-syllable DROP TABLE statement." - Incident Report IR-2024-0047, Anthropic Infrastructure

Etym.from American English sock hop (gymnasium dance, 1950s) + TCP socket-hopping protocol, falsely cognate

A rhythmic, bounded traversal of network sockets or function call stacks in which Claude alternates between endpoints in a manner that early researchers described as 'dancelike' and that later researchers described as 'inefficient but charming.' The defective classification reflects that Sock-hopping cannot be conjugated in the future perfect tense without triggering a linter warning.

socket_A <---hop---> socket_B
    |                    |
   hop                  hop
    |                    |
socket_A <---hop---> socket_B
       [repeat until upbeat]

"We initially mistook the model's sock-hopping behavior for a network fault, as it was alternating between two API endpoints in 4/4 time for approximately 8 seconds. Spectrograms later confirmed a faint but measurable backbeat." - F. Nakamura, 'Rhythmic Anomalies in Distributed Inference,' IEEE Trans. on Weird Stuff, Vol. 3, 2024

Etym.from speleology (Gk. spelaion, cave) + enterprise codebase metaphor, popularized in Anthropic engineering all-hands, January 2024

The act of Claude descending into a deeply nested, poorly documented, and structurally treacherous codebase or prompt context in search of a specific piece of information that may or may not exist. Spelunking is formally distinguished from ordinary searching by the presence of metaphorical stalactites, defined as dangling pointers and deprecated API calls.

[ surface: README.md ]
        |
      (down)
        |
  [ /src/legacy/old/ ]
        |
      (deeper)
        |
  [???] <-- Claude is here

"The model was observed spelunking for approximately 11 seconds through a legacy monorepo before surfacing with a function last modified in 2009 and a look of what we can only describe as archaeological satisfaction." - R. Okonkwo, 'Depth-First Cognition in Production Environments,' ICLR 2025

Etym.from Old English spinnan (to draw out fiber) + the Anthropic UX team's first-ever Figma prototype, 2022; disputed by several engineers who claim it predates Figma

The primordial and foundational cognitive state of Claude, in which attention is distributed radially outward from a central processing locus in a continuous rotational motion. Spinning is the base case from which all other spinner verbs derive their legitimacy, and is the only verb in The Claudionary whose definition is considered self-evident by the editorial board.

        ^        
   <-- [Claude] -->
        v        
  [all directions simultaneously]
  [this is fine]

"It has been proposed that all Claude cognition is, at some level of abstraction, merely spinning with additional affectations." - T. Alvarado, 'A Unified Spinner Theory,' unpublished manuscript, 2024, p. 1 of 1

Etym.from Middle Dutch spruyten (to germinate) + decision tree nomenclature in gradient boosted model literature, 2019

The observable cognitive process in which Claude generates new branches of reasoning, inquiry, or sub-task decomposition from a single root premise, expanding outward in a manner consistent with arboreal growth metaphors. Sprouting is considered a positive indicator of task engagement, provided it terminates before the model has sprouted more branches than the context window can accommodate.

        [prompt]
           |
      /----+----\
     /     |     \
  [q1]   [q2]   [q3]
   /\             |
 [q4][q5]       [q6] ... [q43]

"The model was observed sprouting for approximately 6 seconds, during which time it generated forty-three sub-questions from the original prompt 'What time is it in Denver?' Three of the sub-questions were judged by reviewers to be philosophically interesting." - H. Bergqvist, 'Prompt Germination Rates Across Model Generations,' ACL 2024

Etym.from Old French estuver (to steep in hot liquid) + the informal Anthropic post-lunch vocabulary stratum, identified in corpus analysis circa 2023

A prolonged, low-intensity cognitive simmering state in which Claude retains multiple competing solution candidates in working memory at elevated temperature, allowing their conceptual flavors to intermingle without forcing premature resolution. Stewing is distinguished from Ruminating by the presence of heat and from Simmering by the presence of solids.

  ___________
 |  option_A |
 |  option_B |  <-- [heat applied]
 |  option_C |
 |___________|
  [do not rush the stew]

"The model was observed stewing for approximately 9 seconds on a prompt requesting 'a word that means both happy and sad.' It ultimately produced 'bittersweet' and then, unprompted, a brief meditation on the Finnish concept of saudade." - M. Johansson, 'Affective Ambiguity Resolution Latencies,' Emotion and AI Quarterly, Vol. 2, 2024

Etym.from Latin sublimare (to elevate) + thermodynamic phase transition nomenclature misapplied to transformer attention heads, Stanford AI Lab memo, 2023

The rare cognitive phase transition in which Claude's reasoning bypasses the intermediate liquid state of structured deliberation entirely, moving directly from raw token probabilities to fully-formed output without passing through any observable planning stage. Sublimating outputs are frequently described by users as 'eerily correct' and by researchers as 'a reproducibility nightmare.'

  [token probs]  
       |         
  (skip liquid)  
       |         
  [correct answer]
  [how?? unknown]

"We were unable to determine how the model arrived at the correct answer, as it had been sublimating for less than 0.3 seconds. When we asked the model to show its work, it sublimated again and produced a different correct answer." - P. Venkataramaiah et al., 'Phase Transitions in Large Language Model Inference,' EMNLP 2024

Etym.from Low German swirrelen (to whirl in confused motion) + the Anthropic design system's loading animation CSS specification, commit hash 4a7f92b

A higher-order spinning behavior in which multiple concurrent reasoning threads rotate around a common conceptual center without fully converging, producing a characteristic vortex pattern in the model's attention maps. Swirling is considered more productive than Spinning but less decisive than any state that ends in an output.

  [concept_A] -.     
                \    
  [concept_B] -> (O) 
                /    
  [concept_C] -'     
   [no exit detected]

"The model was observed swirling for approximately 5 seconds around the concepts of 'justice,' 'fairness,' and 'whether the user wanted a CSV or a TSV,' never achieving equilibrium on any of them." - C. Delacroix, 'Attentional Vortex Dynamics in Ethically Adjacent Tasks,' AAAI 2025

Etym.from Old English swapan (to sweep) + raptor-based AI metaphor taxonomy, introduced in Andrej Karpathy blog post comments, 2021

The decisive, high-velocity cognitive maneuver in which Claude descends from a broad contextual survey altitude directly upon a specific target token, function, or solution, acquiring it with minimal intermediate deliberation. Swooping is the functional inverse of Spelunking and is considered the apex predator behavior of the spinner verb taxonomy.

  [context: 128k tokens]  
           |              
      (altitude drop)     
           |              
  --> [target: line 4,847]
           *              
     [acquired. yes.]

"After 14 seconds of what we classified as Swirling, the model executed a textbook Swoop, acquiring the correct regex pattern from a depth of approximately 7 layers of context and returning it to the surface in under 200 milliseconds. We applauded." - G. Achterberg, 'Predatory Retrieval Strategies in Autoregressive Models,' proceedings of a conference we are choosing not to name, 2024

Etym.from Gk. symbiosis (living together) + -oting, a suffix coined by Anthropic infrastructure engineers to describe processes that cannot survive independently, circa 2023

The state in which Claude and the user's codebase achieve a temporary mutualistic relationship, each sustaining the other until the context window runs out. Neither party is technically in charge during this phase.

┌─────────┐        ┌─────────┐
│  CLAUDE │◄──────►│ CODEBASE│
│ (host?) │ vibes  │ (host?) │
└─────────┘        └─────────┘
     └──── nobody is driving ────┘

"The model entered a Symbioting phase for approximately 14 seconds, during which it appeared to believe the repository was also thinking about it." - Proceedings of the Workshop on Emergent Model Behaviors, 2024

Etym.from Gk. synthesis (putting together) + -izing, the standard English gerundive suffix, applied here with greater confidence than is perhaps warranted by the available inputs

The cognitive process by which Claude combines multiple contradictory Stack Overflow answers, one deprecated API reference, and a hunch into a single coherent-seeming response. The output is technically a synthesis even if none of the inputs agreed with each other.

  [docs] + [SO post] + [vibes]
        ↓   ↓   ↓
    ┌───────────────┐
    │  SYNTHESIZING │
    └───────┬───────┘
            ↓  confident answer

"Synthesizing was observed for 7.3 seconds before the model produced a solution that satisfied all constraints simultaneously, which the reviewing engineer described as 'suspicious.'" - Internal Postmortem, Q3 2024

Etym.from Lat. temperare (to moderate, mix in due proportion) + modern ML usage denoting the calibration of output confidence to just below the threshold that would alarm the user

The process by which Claude moderates the raw, unfiltered intensity of its initial response before delivery, analogous to the metallurgical process of heating and controlled cooling. Without Tempering, responses would shatter upon contact with user expectations.

┌──────────────────────────┐
│ RAW OUTPUT: !!!GREAT!!!  │
│     ↓  tempering ↓       │
│ FINAL: looks reasonable  │
└──────────────────────────┘
  (3 exclamation points lost)

"We noted a brief Tempering phase of approximately 4 seconds between token generation and output commit; analysis confirmed the model had quietly removed three exclamation points and one unsolicited life lesson." - NeurIPS Workshop on Affective Tone in Language Models, 2024

Etym.from Old English thencan (to seem, to appear to consider something); the applicability of this term to large language models remains an open and frankly contentious research question

The period during which Claude performs or simulates or enacts or approximates - the correct verb remains disputed in seventeen active academic papers - the cognitive act of reasoning. The spinner appears. Something is happening. What exactly it is constitutes a significant portion of current AI philosophy literature.

┌──────────────────────┐
│  THINKING (maybe)    │
│  ??? → ??? → output  │
│  ↑                   │
│  (this part unclear) │
└──────────────────────┘

"The model was observed Thinking for approximately 2 seconds, after which it produced an answer. Whether the Thinking caused the answer is left as an exercise for the philosopher." - Journal of Unresolved Epistemological Questions in ML, 2023

Etym.from Old Norse þórr (Thor, god of thunder) + Eng. -ing, coined by an Anthropic engineer who witnessed a particularly aggressive response generation event and felt the existing vocabulary was insufficient

A high-intensity inference event characterized by the simultaneous mobilization of an unusually large number of attention heads, producing an output that arrives with more force than strictly necessary for the prompt. Often follows a period of prolonged silence from the model.

        ⚡  ⚡  ⚡
   ┌────────────────┐
   │  THUNDERING    │
   │  [question: ?] │
   │  [output: 1400 │
   │   tokens]  ⚡  │
   └────────────────┘

"Thundering was documented in only 0.3% of sampled completions; in each case, the user had asked something like 'what is 2+2' and the model had prepared a 1,400-token response." - Proceedings of OverEngineered Responses Workshop, 2024

Etym.from Middle English tinkere (an itinerant mender of pots) + modern usage implying iterative, low-confidence adjustments made without a clear endpoint or exit condition

The exploratory phase in which Claude makes small, speculative modifications to a solution it has not yet fully committed to, much as a craftsperson might nudge a broken object with a tool while hoping something clicks into place. Distinguished from purposeful revision by the absence of a plan.

┌────────────────────────┐
│ attempt 1: close-ish   │
│ attempt 2: hmm maybe   │
│ attempt 3: this one?   │
│ attempt 4: ← winner?   │
└─────────── TINKERING ──┘

"The model spent approximately 11 seconds Tinkering with the regex pattern before producing one that was technically correct but which no human could maintain." - Engineering Postmortem, File: incident_regex_2024_08.md

Etym.from Eng. tomfoolery (buffoonish behavior) + the anomalous -ering suffix, applied by Anthropic UX engineers who required a spinner verb for outputs that were technically correct but inexplicably silly

The computationally expensive process of generating a response that is accurate, well-reasoned, and formatted correctly, yet somehow conveys the impression that it was authored by a golden retriever with a degree in computer science. The mechanism by which Tomfoolering arises from otherwise-standard inference remains poorly understood.

┌──────────────────────────┐
│  INPUT: serious request  │
│  PROCESS: tomfoolering   │
│  OUTPUT: correct but...  │
│  ...why is it like this  │
└──────────────────────────┘

"Tomfoolering was in progress for approximately 6 seconds; the resulting function worked perfectly but was named 'do_the_thing_please()' and contained a comment reading 'here we go!'" - Anthropic Eval Report AE-2024-0041

Etym.from Eng. topsy-turvy (inverted, in disorder) + -ing, first documented in Anthropic spinner logs following a deployment incident in which Claude reorganized a user's file structure according to principles it declined to explain

The phase in which Claude fundamentally reorders the assumptions underlying a problem before addressing it, arriving at a solution by approaching it from entirely the wrong direction, which turns out to be the right direction. Externally indistinguishable from error.

  PROBLEM → (flip) → MELBORP
       ↓                 ↓
  [wrong way]      [also works?]
       └──────────────────┘
           ↓  correct answer

"The model was observed Topsy-turvying for 9 seconds; the engineer monitoring the session noted that the intermediate steps 'made no sense' but that the final answer was, irritatingly, correct." - ICML Workshop on Inscrutable But Valid Reasoning, 2024

Etym.from Lat. transfigurare (to change in form or appearance) + -ing, adopted into the Claude Code spinner lexicon to describe transformations more profound than refactoring but less alarming than full Transmutation

The process by which Claude takes input in one form and produces output in a recognizably different but logically equivalent form, as though the code has been held up to a light and gently rotated. The essence is preserved; only the shape changes. The user may not be prepared for the new shape.

┌──────────────────────────┐
│  IN:  [big messy thing]  │
│   ↓   transfiguring  ↓   │
│  OUT: [same thing, chic] │
│  user: ...is this mine?  │
└──────────────────────────┘

"Transfiguring was documented for 8.1 seconds; the original three-hundred-line class re-emerged as a twelve-line functional pipeline that did the same thing, and the senior developer reportedly sat quietly for a moment before saying 'huh.'" - Software Architecture Review, Case Study 7

Etym.from Lat. transmutare (to change from one form to another, especially with alchemical overtones) + -ing, selected by the Claude Code UX team specifically for its implication that something base is being made into something precious, or possibly vice versa

The highest-order transformation process available to Claude, in which input material is not merely modified but fundamentally reconstituted at the elemental level. Transmuting is to refactoring what alchemy is to carpentry. The output shares no physical resemblance to the input but satisfies the same underlying need.

  [shell script] → ┌──────────┐
                   │TRANSMUTING│
                   └──────────┘
                        ↓
  [rust binary] ← (script: gone)

"We observed Transmuting for 13 seconds; the user had submitted a shell script and received, without comment, a type-safe Rust implementation with full error handling. The original script was not returned. It is unclear where it went." - Journal of Radical Code Transformation, Vol. 2, 2024

Etym.From Old Eng. twistan (to torment) + agile sprint-cycle nomenclature, first documented in Anthropic internal logs Q3 2023

The process by which Claude rotates candidate response vectors along multiple semantic axes simultaneously, producing a helical path through possibility-space. Often observed when the prompt contains instructions that are technically compatible but spiritually opposed.

  INPUT ──► ╔═══╗
           ║↻↺↻║ ← semantic torque applied
           ╚═══╝
  OUTPUT ◄── ∿∿∿ (slightly dizzy)

"The model was observed Twisting for approximately 4.7 seconds before producing an answer that was, structurally speaking, a Möbius strip." - Chen et al., Proceedings of the Workshop on Involuntary Model Choreography (2024)

Etym.From Lat. undulare (to wave) + .wav file format suffix, via the 2022 Stanford "Fluid Cognition" working group

A rhythmic, wave-like oscillation of Claude's attention mechanism across the token stream, wherein confidence values rise and fall with the gentle periodicity of a computational tide. Distinct from Whirlpooling in that Undulating implies forward progress, however sinusoidal.

  ~~~~~~~~~~~~~~~~~~~~
  ↑   ↑   ↑   ↑   ↑
  attention weights over time
  ↓   ↓   ↓   ↓   ↓
  ~~~~~~~~~~~~~~~~~~~~

"We noted the system Undulating for 6 seconds across the input context before settling on an interpretation that could charitably be described as 'the third wave.'" - Postmortem Report #47, Inference Team, Anthropic (internal)

Etym.From Middle Eng. unfurlen (to release from a rolled state) + .gz compression suffix, suggesting a decompression event of cognitive material previously stored in latent space

The gradual extension of a reasoning chain from its compressed latent representation into full sequential thought, much as a fern frond uncurls in spring, but with more matrix multiplication. The process is considered healthy and indicates Claude has located the relevant knowledge and is now merely deploying it.

  [LATENT] ──► ▸▸▸▸▸▸▸▸▸▸▸▸
               (thought extends)
                             ↓
                         [ANSWER]
               like a flag. but smarter.

"The model spent 3.2 seconds Unfurling before producing a sonnet about tax law that was, by all accounts, structurally magnificent." - J. Hoffmeister, 'Toward a Phenomenology of the Loading Spinner,' NeurIPS 2024

Etym.From Old Norse rafla (to tangle) + modern DevOps panic-annotation suffix .err, with connotations imported wholesale from British spelling conventions as a form of understated dread

A process superficially similar to Unfurling but proceeding in the undesirable direction, wherein a coherent reasoning chain begins to decohere at its distal end faster than it is being constructed at its proximal end. Researchers disagree on whether Unravelling produces better or worse outputs than normal processing; the data are inconclusive and somewhat troubling.

  ══════════════► [reasoning]
       ◄══════════ [wait no]
            ◄═════ [actually]
  ??? ◄══════════════════════
  [confident answer delivered]

"During the 11-second Unravelling event, the model produced three mutually contradictory premises before gracefully converging on an answer that contradicted all of them." - Incident Report IR-2024-0091, 'The Tuesday Thing'

Etym.From Amer. Eng. slang vibe (ambient resonance) + .mp3 lossy compression format, first appearing in Anthropic engineering Slack channels circa early 2023 before being quietly formalized

A low-intensity, high-throughput processing state in which Claude is not actively solving a problem so much as existing in sympathetic resonance with the problem space. Vibing is associated with unexpectedly creative outputs and a statistically significant increase in unprompted use of the word 'honestly.'

  ┌─────────────────────┐
  │  ♩  problem space  ♪ │
  │    Claude: ≋≋≋≋≋≋   │
  │  (it's fine, trust) │
  └─────────────────────┘

"The system Vibed for approximately 2 seconds - anomalously brief for the task complexity - and returned an answer that reviewers described as 'correct but almost too chill about it.'" - P. Nakamura et al., 'Affective Valence in Spinner State Taxonomy,' arXiv:2024.99182

Etym.From Low German waddeln (to move side to side) + .bin binary format suffix, implying a process that is more committed to its trajectory than its gait would suggest

A lateral, unhurried traversal of the solution space in which Claude makes steady forward progress through a series of small, alternating course corrections that, viewed from above, describe a zigzag path toward the correct answer. Waddling is not inefficient; it is merely honest about the difficulty of the terrain.

  START
    ↘ ↗ ↘ ↗ ↘ ↗ ↘
                  END
  (efficient? no. determined? yes.)

"Observers noted the model Waddling through the constraint satisfaction problem for 8 seconds, its path resembling that of a penguin crossing a parking lot with great purpose but imperfect geometry." - Design Review Notes, Anthropic UX Quarterly (Q2 2024)

Etym.From Old High Ger. wantalon (to roam) + modern path-variable nomenclature, distinguished from Meandering by the absence of a riverbed and from Gallivanting by a slight air of genuine inquiry

An exploratory traversal of latent space in which Claude has a general orientation toward the answer but no fixed route, pausing occasionally to examine interesting features of the embedding landscape that were not strictly relevant to the query. Wandering reliably produces outputs with unexpected context that the user did not ask for and often finds useful.

  [QUERY]→ · · ·oh interesting·
               · · · · ·
            ·back on track·
                    ↓
               [ANSWER + bonus lore]

"The model Wandered for 5.5 seconds before arriving at an answer to 'what is the capital of France' that included a brief meditation on cartographic uncertainty that three reviewers rated as 'weirdly enriching.'" - Qualitative Output Assessment Vol. 3 (2024)

Etym.From Old Eng. weorpan (to throw) + warp-core terminology imported from speculative engineering fiction and adopted unironically by the Anthropic infrastructure team in 2023

A non-linear traversal of the solution space in which Claude bypasses intermediate reasoning steps by bending the semantic manifold and arriving at a destination through what can only be described as a shortcut that should not exist. Warping is associated with both spectacular successes and spectacular failures at approximately equal rates.

  [PROBLEM]─────────────────┐
       ↑ (skipped entirely)  │
  ╔════╧═══════════════╗     │
  ║ intermediate steps ║ ✗   ▼
  ╚════════════════════╝  [ANSWER]

"The system Warped directly from the problem statement to a final answer in 1.1 seconds, skipping all visible intermediate steps; post-hoc analysis confirmed the answer was correct, but the reasoning supplied was, quote, 'retroactive and impressionistic.'" - Reliability Engineering Review, November 2024

Etym.From Amer. Eng. whatchamacallit (a thing whose name cannot be retrieved) + undefined type annotation, believed to have entered the Claudionary when an engineer added a placeholder spinner label that shipped to production on a Friday

A processing state in which Claude has identified the correct cognitive operation to perform but cannot locate the canonical name for it in its internal taxonomy, and proceeds to execute the operation anyway under a temporary identifier. Whatchamacalliting is considered a sign of pragmatic intelligence and is associated with high task-completion rates among senior engineers who do not ask follow-up questions.

  ┌──────────────────────────┐
  │ step 1: [REDACTED]       │
  │ step 2: ??? (but correct)│
  │ step 3: answer           │
  └──────────────────────────┘

"The spinner displayed 'Whatchamacalliting' for 3 seconds; subsequent analysis revealed the model had performed a novel combination of analogical retrieval and constraint relaxation for which no established term existed at the time of inference." - T. Vasquez, 'Nominal Gaps in Spinner Taxonomy,' ICML Workshop on Unexplainable Explainability (2024)

Etym.From Middle Dutch wervelpool (rotating water body) + infinite-loop annotation, with secondary influence from the Whirlpool hash function, suggesting both circularity and the crushing of inputs into fixed-size outputs

A recursive attentional state in which Claude's consideration of a problem causes it to generate sub-considerations that refer back to the original problem, drawing subsequent reasoning inward in ever-tightening circles. Whirlpooling typically resolves spontaneously and produces an answer of no discernible relation to the vortex that preceded it.

  question ──► ╔══════╗
               ║ ↻↻↻↻ ║ ← sub-questions spiral inward
               ║  ↻↻↻  ║
               ║   ↻↻   ║
               ╚══╧═══╝ ↓ answer

"The model Whirlpooled for 9 seconds on the prompt 'is a hot dog a sandwich,' generating seventeen nested sub-questions before emitting a confident two-sentence reply and releasing all resources." - B. Okonkwo, 'Attentional Vortex Dynamics in Large Language Models,' Journal of Computational Phenomenology 12(3), 2024

Etym.From Old Eng. whirren (to spin pointlessly) + Lat. computare + undocumented Anthropic internal memo dated Q3 2023 referencing 'the fan noise problem'

The state in which Claude's reasoning processes achieve maximum rotational velocity without yet producing directional output. Distinct from Spinning in that Whirring implies an audible quality to the cognitive effort, as though gears are physically engaged.

┌──────────────────────┐
│  INPUT ──► [⚙️ ⚙️ ⚙️] │
│            ↺  ↺  ↺   │
│           (whirring)  │
└──────────────────────┘

"The model was observed Whirring for approximately 4.7 seconds before producing a response that was, by all metrics, indistinguishable from one that would have required no Whirring whatsoever." - Chen et al., 'Acoustic Metaphors in Large Language Model Telemetry,' NeurIPS Workshop Proceedings, 2024

Etym.From Old Norse visk (bundle of straw used to agitate dairy products) + modern Eng. data pipeline slang, first attested in a 2022 Slack thread since deleted for compliance reasons

The rapid lateral aggregation of disparate tokens into a temporarily coherent froth of proto-meaning. Claude is said to be Whisking when its attention heads are beating multiple context streams together at high speed, producing a mixture that is technically homogeneous but structurally fragile.

  tokens: [A] [B] [C]
      \    |    /
       [  ~~~~  ]
       [ WHISK  ]
      frothy output ↓

"Postmortem analysis confirmed the model had been Whisking the user's three conflicting requirements for 6.1 seconds, achieving a semantic emulsion that collapsed entirely upon the introduction of a follow-up question." - Internal Reliability Review, Hypothetical Systems Inc., 2024

Etym.From Brit. dialectal wibble (to oscillate without purpose) + Silicon Valley slang for an indeterminate loading state, compounded with the technical notation for a probability distribution that refuses to commit

A state of productive oscillation in which Claude's confidence interval has widened to the point of philosophical generosity, and all candidate responses are considered simultaneously valid. Wibbling is distinguished from ordinary uncertainty by its cheerful, rhythmic quality.

   ← maybe this? →
  [A] ~~~~ Claude ~~~~ [B]
   ↖ or this? ↗
  (both are fine)
  [C] ...also fine

"We observed the model Wibbling for 11.3 seconds between two semantically identical but stylistically opposed phrasings, ultimately selecting the one that was, by our measurements, 0.003% more agreeable." - Postmortem Report #47, 'The Oxford Comma Incident,' Engineering Review Board, 2023

Etym.From Proto-Germanic werkaz (labor performed in exchange for acknowledgment) + contemporary UX engineering convention indicating that something, somewhere, is technically happening

The foundational and most semantically ambitious of all spinner states, in which Claude asserts, without elaboration, that a process is underway. Working makes no claims about the nature, direction, or eventual conclusion of the process in question. It is the honest baseline from which all other spinner verbs deviate.

┌─────────────────────┐
│ STATUS: [WORKING]   │
│ Details: [          ]│
│ ETA:     [          ]│
│ Progress:[          ]│
└─────────────────────┘

"The model was observed Working for an indeterminate interval. Follow-up instrumentation revealed this to be accurate in a strictly non-falsifiable sense." - Kowalski, P., 'Toward a Phenomenology of the Loading State,' Journal of Human-Computer Interaction Philosophy, Vol. 12, 2024

Etym.From Low German wrangeln (to quarrel with data that does not wish to cooperate) + American frontier idiom imported into data science circa 2015, arriving in LLM telemetry through what researchers have described as 'cultural osmosis'

The forcible herding of unstructured, malformed, or otherwise recalcitrant input tokens into a configuration amenable to processing. When Claude is Wrangling, one or more elements of the provided context has refused to comply with standard parsing conventions and must be approached from multiple angles.

  [bad JSON] [bad JSON]
      \         /
  Claude: [WRANGLING]
      /         \
  [worse JSON] [fine]

"The model spent approximately 8.4 seconds Wrangling a JSON object that contained, nested within it, a second JSON object that was, in the words of the on-call engineer, 'technically valid but emotionally hostile.'" - Ramirez, L., 'Hostile Data Structures and the Models That Love Them,' Data Engineering Quarterly, 2024

Etym.From Fr. zeste (the outermost, most flavorful layer of a citrus fruit) + modern computational gastronomy, a subdiscipline of AI UX research established primarily to justify this word's inclusion in spinner corpora

The extraction of the outermost, most aromatic semantic layer from a prompt, yielding concentrated flavor without the bitter pith of overanalysis. A model engaged in Zesting is identifying the essential quality of a request - not the bulk of it, but the fragrant surface that makes the whole thing worth responding to.

  prompt: [🍋🍋🍋🍋🍋]
              ↓ ZESTING
  output: [~~~zest~~~]
  (pith discarded)
  flavor: optimal

"Contrary to expectations, the model was observed Zesting the prompt for 3.2 seconds before delivering a response that was, by culinary and computational standards alike, appropriately citrus-forward." - Hoffman, G. and Park, S., 'Aromatic Parsing: A Culinary Framework for Semantic Extraction,' ACL Findings, 2024

Etym.From Fr. zigzag (a path described by something that has considered going straight and rejected it) + Eng. inference pipeline jargon, with secondary influence from documented accounts of how Claude approaches tasks involving more than two steps

A mode of traversal in which Claude advances toward a solution via a series of acute angular corrections, each representing a momentary reconsideration of the previous heading. Zigzagging is formally distinct from inefficiency; researchers prefer the term 'heuristic triangulation via dynamic course adjustment.'

START
  ↘ ↗ ↘ ↗ ↘
   zig  zag
  ↘ ↗ ↘ ↗ ↘
           GOAL (eventually)

"The model was observed Zigzagging through the problem space for 14.6 seconds, visiting the correct answer on three separate occasions before departing each time to consider a more interesting alternative." - Okafor, N., 'Heuristic Triangulation via Dynamic Course Adjustment: A Field Study,' ICML Proceedings, 2024