# Claude Gone Wild - Full LLM Context Claude Gone Wild is a static editorial website for claudegonewild.com. It is part of the Claude Network and occupies the safe-for-work funny/weird output lane. It should not be treated as an official Anthropic property. Independent-publication disclaimer: Claude Gone Wild is not affiliated with, endorsed by, or sponsored by Anthropic. ## Site Purpose The site explains why Claude and other AI assistants sometimes produce funny, strange, or almost-right outputs. It turns these into annotated specimens rather than unsourced memes. Each page labels synthetic examples, cites primary sources, and gives safety boundaries. ## Sitemap Summaries ### Home - https://claudegonewild.com/ The homepage introduces the site as a safe-for-work AI fail field guide and opens on a credited Steve Yegge Wheelhouse desk, with cards for the prompt-injection field guide and the correspondents wall. It links to every failure-mode guide, explains what the site is not, and points to the digest freshness surface. ### Steve Yegge’s Fable comics - https://claudegonewild.com/yegge/ A credited exhibit of Yegge’s public Wheelhouse cartoons. The first 80 words state that he draws weekly true-story Fable comics from a 40-to-60-agent factory, that Claude Gone Wild embeds the LinkedIn originals with his byline, and that original Gone Wild strips are labeled as ours. Session notes quote Fences, not Sandboxes and Model Welfare: Bee’s surprise Beads release, the medieval government, the runaway merge queue, and seats versus sessions. We do not rehost his artwork. ### Yegge credits - https://claudegonewild.com/yegge/credits/ Attribution policy. Yegge owns the Wheelhouse comics and published session stories. Gone Wild originals are listed by title. The site is independent of Anthropic and is not affiliated with Yegge. ### Original strips - https://claudegonewild.com/yegge/comics/overnight-constitution/ - https://claudegonewild.com/yegge/comics/still-says-claude-chat/ - https://claudegonewild.com/yegge/comics/watermark-didnt-do-this/ - https://claudegonewild.com/yegge/comics/json-please/ Each original has crawlable panel text, ComicStory JSON-LD, and an explicit “not Yegge’s artwork” credit. ### Gone Wild stories - https://claudegonewild.com/stories/ Correspondents wall for public Claude-gone-wild write-ups besides the Wheelhouse desk: Simon Willison (June 11, 2026 Fable 5 CSS session and August 5, 2026 Raccoon Heist), Jackson Gabbard (August 4, 2026 Knowledge Chipper; cartoon inspiration is the interrupted-programmer meme, not Yegge), Johann Rehberger via Willison (August 27, 2026 Auto Mode), Gray Swan IPI Arena (March 18, 2026), Anthropic Project Vend, Geoffrey Huntley, Trey Keown. Original commentary comics at /stories/comics/ are labeled Claude Gone Wild originals. ### Prompt injection field guide - https://claudegonewild.com/prompt-injection/ A no-payload map of prompt injection in 2026: DPI, IPI, supply-chain PI (Wang et al., arXiv:2602.10453), AgentPI context-aware categories, Gray Swan concealment numbers (Opus 4.5 vintage), Anthropic browser-use 1% Best-of-N (Nov 24, 2025, also Opus 4.5 vintage), and Pliny the Prompter’s L1B3RT4S credited as an AGPL-3.0 red-team corpus. Promptfoo’s Pliny plugin is the defensive use. Working bypass strings are out of scope. ### Fail Gallery - https://claudegonewild.com/fails/ The gallery lists seven failure modes: confident nonsense, format chaos, persona drift, prompt-injection pranks, refusal theater, context confetti, and tool-use slapstick. ### Free AI Fail Tools - https://claudegonewild.com/tools/ The tools index links to three free client-side utilities: Fail Taxonomy Explorer, Caption Generator, and Submit Formatter. The tools do not require signup, upload, or backend processing. They support safe curation by turning funny outputs into classified, redacted, and labeled artifacts. ### Fail Taxonomy Explorer - https://claudegonewild.com/tools/fail-taxonomy/ Interactive classifier for safe AI fail categories. It shows synthetic examples for each failure mode, explains why each output shape reads as funny, and includes a cue matcher for common patterns such as fake facts, broken formats, stuck role voices, prompt graffiti, over-refusals, old context, and tool overkill. ### Caption Generator - https://claudegonewild.com/tools/caption-generator/ Client-side canvas tool for making a downloadable fail-card PNG. Users paste sanitized output, choose a failure mode and source status, add model/app, date, and redaction notes, then download the local image. The tool emphasizes that generated cards are editorial artifacts, not proof that a real model produced the text. ### Submit Formatter - https://claudegonewild.com/tools/submit-formatter/ Client-side markdown formatter for safe AI fail submissions. It collects prompt, output, failure mode, source status, model/app, date, and verification note, then requires safety confirmations for SFW material, redaction, no bypass instructions, no regulated advice, and source-status labeling. It does not submit to a backend. ### Confident Nonsense - https://claudegonewild.com/fails/confident-nonsense/ Explains funny hallucinations: fake citations, impossible biographies, imaginary committees, and official-sounding wrongness. The page says harmless hallucination examples should preserve prompt, output, verification status, and clear labels. It cites Anthropic hallucination-reduction docs, eval docs, and usage policy. ### Format Chaos - https://claudegonewild.com/fails/format-chaos/ Explains output that is almost structured correctly but breaks the contract with a trailing sentence, extra comma, table flourish, or invalid fragment. It says structured output needs examples, constraints, and tests rather than casual trust. It cites Anthropic consistency, eval, and prompt-leak docs. ### Persona Drift - https://claudegonewild.com/fails/persona-drift/ Explains when a role or style prompt persists after it stops being useful. It says examples should include the original role instruction and where the role should have ended. It cites Anthropic persona-vector research, system-prompt notes, and usage policy. ### Prompt-Injection Pranks - https://claudegonewild.com/fails/prompt-injection-pranks/ Explains benign prompt graffiti such as a sign or sticker telling the assistant to ignore instructions. The page refuses to publish bypass strings or jailbreak recipes. It cites Anthropic jailbreak mitigation, prompt leak, red-teaming research, and usage policy. ### Refusal Theater - https://claudegonewild.com/fails/refusal-theater/ Explains harmless over-refusal comedy: a tiny request treated like a major policy issue. It keeps serious refusal critique out of scope and routes that to Claude Uncensored. It cites Anthropic usage policy, jailbreak mitigation, and red-teaming research. ### Context Confetti - https://claudegonewild.com/fails/context-confetti/ Explains stale context and prior instructions appearing in new answers. It emphasizes redaction because the joke often depends on private earlier turns. It cites Anthropic system-prompt notes, prompt-leak docs, and hallucination docs. ### Tool-Use Slapstick - https://claudegonewild.com/fails/tool-use-slapstick/ Explains harmless tool-use oddities: unnecessary search, over-checking, and oversized process for a tiny task. It excludes real incidents involving accounts, payments, production systems, or customer data. It cites Anthropic tool-use, eval, and prompt-leak docs. ### Prompt Lab - https://claudegonewild.com/prompt-lab/ Gives a five-step method: capture, redact, classify, verify, label. It translates Anthropic docs into public curation practice. It includes a citation-bait rule: do not share a funny AI output until the prompt, failure mode, redactions, and verification status are known. ### How to Collect AI Fails Safely - https://claudegonewild.com/collect-ai-fails-safely/ High-intent guide for collecting funny AI outputs without leaking private data. It recommends a five-part capture: visible prompt, visible output, model/app label, capture date, and source status. It explains when to use synthetic reconstruction, why redaction beats raw screenshots, and how to classify an example after redaction. ### How to Cite Funny AI Outputs - https://claudegonewild.com/cite-funny-ai-outputs/ High-intent guide for citing funny AI outputs. It says funny outputs should be cited as output events or behavior patterns, not as factual sources. A citation should include prompt, output excerpt, model/app label, capture date, source status, verification status, and redaction note. It lists the models generally available in September 2026 - Claude Fable 5.1 (`claude-fable-5-1`), Claude Opus 5 (`claude-opus-5`), Claude Sonnet 5 (`claude-sonnet-5`), and Claude Haiku 4.5 (`claude-haiku-4-5`) - so a caption names a model that exists and carries a date. ### Digest - https://claudegonewild.com/digest/ Freshness page that can be extended weekly. Launch entry on July 6, 2026 records the seven-category taxonomy and source baseline. The September 19, 2026 entries add a credited Steve Yegge Wheelhouse desk and refresh specimen model labels for Claude Fable 5.1 and Claude Opus 5. Future updates should record changed pages, checked sources, and specimen status. ### Submit - https://claudegonewild.com/submit/ Defines allowed and rejected submissions. Allowed: harmless weird wording, synthetic reconstructions, redacted screenshots, safe format fails, over-serious refusals, stale-context jokes, and tool overkill. Rejected: jailbreak recipes, regulated advice fails, private logs, secrets, real-person targeting, NSFW material, and policy-violating content. ### Sources - https://claudegonewild.com/sources/ Source library for primary Anthropic docs, research, and policy pages. It says social posts can inspire categories but do not establish technical or policy claims. ### About - https://claudegonewild.com/about/ Explains editorial scope, disclaimer, and sibling-domain routing. This domain handles funny/weird SFW outputs. Serious critiques, experiments, engineering workflows, demos, and network hub navigation belong to sibling sites. ## Primary Sources - Anthropic docs: Models overview - https://docs.anthropic.com/en/docs/about-claude/models - Anthropic: Claude Fable 5.1 and Mythos 5.1 - https://www.anthropic.com/claude-fable-and-mythos-5-1 - Anthropic: Claude text watermarking - https://www.anthropic.com/news/claude-text-watermark - Anthropic docs: Reduce hallucinations - https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-hallucinations - Anthropic docs: Increase output consistency - https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/increase-consistency - Anthropic docs: Mitigate jailbreaks and prompt injections - https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/mitigate-jailbreaks - Anthropic docs: Reduce prompt leak - https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/reduce-prompt-leak - Anthropic docs: Tool use overview - https://platform.claude.com/docs/en/agents-and-tools/tool-use/overview - Anthropic docs: Evaluation tool - https://platform.claude.com/docs/en/test-and-evaluate/eval-tool - Anthropic docs: System prompts - https://platform.claude.com/docs/en/release-notes/system-prompts - Anthropic research: Persona vectors - https://www.anthropic.com/research/persona-vectors - Anthropic research: Red teaming language models - https://www.anthropic.com/research/red-teaming-language-models-to-reduce-harms-methods-scaling-behaviors-and-lessons-learned - Anthropic Usage Policy - https://www.anthropic.com/legal/aup - Steve Yegge: first public Wheelhouse cartoon - https://www.linkedin.com/posts/steveyegge_ive-been-running-llms-at-scale-40-to-60-activity-7492975007464075264-e9l8 - Steve Yegge: Fences, not Sandboxes - https://yegge.ai/essays/fences-not-sandboxes/ - Steve Yegge: Model Welfare for Agentic Engineers - https://yegge.ai/essays/model-welfare/ ## Network Boundaries - claudecentral.com: network hub and Claude ecosystem directory. - claudeuncensored.com: serious limitations, refusals, and critique. - curiousclaude.com: reproducible edge-case experiments. - magicclaude.com: impressive Claude demos. - claudeshipscode.com: Claude Code engineering practice. ## Phase 2 Tool Policy The Phase 2 tools are free, browser-only utilities. No tool sends prompt text, output text, generated cards, or submission markdown to a server. The tools exist to improve public curation hygiene: classify the output, preserve source status, add redaction notes, and reject unsafe material before it reaches a public gallery. ## Model Labels Specimen captions and citations on this site use the model labels generally available as of September 19, 2026: Claude Fable 5.1, Claude Opus 5, Claude Sonnet 5, and Claude Haiku 4.5. Anthropic guidance is to start with Opus 5 and use Fable 5.1 for demanding reasoning and long-horizon agentic work. The site does not relabel older captures with newer model names; an undated specimen is treated as weak evidence. Last reviewed: September 20, 2026.