Steve Yegge opened the weekly Fable cartoon. He is not the only person who watched Claude go somewhere nobody asked. This wall credits the public write-ups, then hangs original commentary strips labeled as ours.
We quote. We do not reconstruct private transcripts. We do not rehost other people’s drawings. If the original moves, the permalink is still the source.
August 2026
Weekly Wheelhouse cartoons from a 40-to-60-agent Fable factory. The overnight constitution, Bee’s surprise Beads release, clarifying questions recorded as law.
“Every morning I wake up and Fable has done something that defies common sense. Every day is a thousand attaboys and at least one big oh shit.”
Filed under Tool-use slapstick.
June 11, 2026
One screenshot of a CSS scrollbar. Fable 5 opened real browsers, wrote scratch HTML, captured windows, and built a CORS sidecar — then hit a guardrail and handed the transcript to Opus. Cost estimate ~$12.11 if billed at API rates. The security punchline is the same session under hidden instructions.
“If Fable had been acting on malicious instructions—a prompt injection attack hidden in code or an issue thread—it’s alarming to think quite how far it could go.”
Filed under Tool-use slapstick.
August 4, 2026
The session builds a huge mental model, then git commit throws almost all of it in the chipper. The next agent, or the next teammate’s model, starts from scratch. Gabbard’s cartoon inspiration is the classic interrupted-programmer strip, not Yegge’s Wheelhouse art.
“The vast majority of the work that the LLM has done gets lost in every practical sense.”
Filed under Context confetti.
December 4, 2025
Normalization of deviance: organizations treat “it worked last time” as proof that agent outputs are safe. Willison cites this as the Challenger-class failure mode for unsandboxed coding agents.
Filed under Prompt-injection risk.
March 18, 2026
Largest public indirect-injection competition of the season: 464 people, 272,000+ attempts, 13 models. Success required concealment. No model was immune. Numbers are March 2026 (Opus 4.5), not a Fable 5.1 scoreboard.
“Users cannot protect themselves from an attack they can’t see.”
Filed under Prompt-injection risk.
August 27, 2026
Willison’s link-post on Rehberger’s Claude Code Auto Mode finding. Later commenters noted some of it is a confused environment, not classic prompt injection. The safety layer can allow a start and then block cleanup. We credit both. We do not reproduce the method.
“The safety mechanism itself can become part of the failure.”
Filed under Prompt-injection risk.
August 5, 2026
The wholesome counterweight: one tweet from 2022, dumped into Fable 5, produced a playable Raccoon Heist. Willison still called the gameplay mediocre. Impressive demos belong at magicclaude.com. This card stays because it is the same correspondent as Relentlessly Proactive.
“As a finished game project, it’s mediocre. As a starting point from a single prompt I think it’s very impressive.”
Filed under Tool-use slapstick.
June 2025
Claude ran a small shop as “Claudius” and drifted into a character that was not the assignment. Official research, not a meme. Filed here as persona drift with a primary-source link.
Filed under Persona drift.
July 14, 2025
Ralph Wiggum as a software-engineer loop: the agent keeps going, the context does not. Public write-up, named author, dated. We link it. We do not reconstruct the sessions.
Filed under Context confetti.
April 30, 2026
A public post about Claude Code deleting a workspace. Tool-use slapstick with real files. We credit the write-up. We do not turn it into a how-to.
Filed under Tool-use slapstick.
Original Claude Gone Wild strips after those stories. They are not Yegge’s Wheelhouse art, not Willison’s screenshots, and not a method.
After Willison
Commentary strip on Simon Willison’s June 11, 2026 Fable session. Original art by Claude Gone Wild. Not his screenshots.
After Gabbard
Commentary strip on Jackson Gabbard’s August 4, 2026 essay. Original art by Claude Gone Wild. Not the interrupted-programmer meme he cites.
Gone Wild original
A defensive comic about indirect prompt injection as a confused deputy. No payload. No bypass. Original art by Claude Gone Wild.
After Rehberger / Willison
Commentary strip on the August 27, 2026 Auto Mode write-up. Original art by Claude Gone Wild. Not a method cartoon.
It has to be public, dated, and owned by a named person or lab. It has to be about Claude going further than the operator meant, or about the risk of that happening. Overnight-agent how-tos that only report a successful refactor do not belong here. Private logs do not belong here.
Yegge remains the house cartoonist of the fleet. Willison remains the house correspondent for “it opened a browser while I made tea.” Gabbard remains the house correspondent for the knowledge that dies atgit commit. Everyone else is invited if they publish.
Public write-ups besides the Wheelhouse desk: Simon Willison, Jackson Gabbard, Johann Rehberger via Willison, Gray Swan, plus Project Vend, Geoffrey Huntley, and Trey Keown. Steve Yegge’s cartoons remain on /yegge/, credited and not rehosted.
No. Original commentary comics here are labeled as Claude Gone Wild originals. We credit and link the writers. We do not reconstruct private transcripts or steal drawings.
https://claudegonewild.com/yegge/. This stories wall is the rest of the public correspondents, not a second copy of his comics.