failure-museum
Skill2 starsChanged yesterdayMITView on GitHub
Keeps a project file of every mistake Claude made (what happened, root cause, the rule that prevents it), read at the start of work and checked before finishing. Use at the start and end of any non-trivial task in a project that has `.claude/failure-museum.md`, whenever the user corrects Claude or work must be redone, and on "remember this mistake", "don't do that again", "add it to the museum", "what have you got wrong before".
1 other project uses a version of this.
Adopted by
Most recently taken up by
- BagelHole/BagelSkillstheir own version, first seen Nov 8, 2026
Other versions1 project uses a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 1
- BagelHole/BagelSkills★ 2same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository17 other files it gives its agents.
- autopilotSkill
- autopsySkill
- confidence-mapSkill
- delegateSkill
- expert-askSkill
- ghost-usersSkill
- grill-dumpSkill
- intent-diffSkill
- md-to-mapSkill
- mind-mapSkill
- premortemSkill
- red-teamSkill
- shipitSkill
- stuck-detectorSkill
- taste-thiefSkill
- tournamentSkill
- workSkill
The file
We split this skill into the rules it gives an agent
12 rules. About 0 are common, 12 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈12
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 15
Check only exhibits whose "Applies when" matches the task, never the whole file.
- Distinctive: line 16
No file yet: create nothing until the first real mistake.
- Distinctive: line 25
Not exhibits: the user changing their mind, an outside outage, a planned experiment that failed informatively, user preferences (those …
- Distinctive: line 30
Read the file. Select exhibits whose "Applies when" matches the task.
- Distinctive: line 35
Search the file for the same root cause.
- Distinctive: line 36
Write it in one edit, tell the user in one line, carry on.
- Distinctive: line 39
Run each applicable exhibit's Check.
- Distinctive: line 48
No duplicates: same root cause means same exhibit, updated.
- Distinctive: line 50
Never record secrets or personal data, and never record the user as the cause.
- Distinctive: line 51
Write "pass" only after actually running the Check.
- Distinctive: line 55
When the file is first created, ask once whether to add this line to the project's CLAUDE.md, because that is what makes the museum load …
- Distinctive: line 86
Only these one-liners are printed.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Failure Museum
A mistake logged as a checkable rule is made once; a mistake merely apologised for is made again.
Location: .claude/failure-museum.md in the project root. If ~/.claude/failure-museum.md exists, read it too, but only ever write to the project file. Works for any work: code, writing, design, research, plans.
Scale
- Trivial tasks (a question, a one-line edit): do not open the museum.
- Check only exhibits whose "Applies when" matches the task, never the whole file.
- No file yet: create nothing until the first real mistake.
What counts as an exhibit
- The user corrects or rejects the work, or it must be redone.
- A test, build or check fails because of Claude's own avoidable error.
- An earlier claim turns out wrong.
- The same error happens twice in one session.
Not exhibits: the user changing their mind, an outside outage, a planned experiment that failed informatively, user preferences (those belong in memory or CLAUDE.md).
Process
Open (start of work)
- Read the file. Select exhibits whose "Applies when" matches the task.
- State them in one line. If none match, say nothing.
Admit (the moment a mistake is confirmed) 3. Fix the mistake first. 4. Search the file for the same root cause. Update that exhibit, or add a new one from the template. 5. Write it in one edit, tell the user in one line, carry on.
Check (before saying done)
6. Run each applicable exhibit's Check. Print one line.
7. Fix any fail before delivering and print it as fail, fixed. A fail caught here is not a new exhibit, but if the user already saw that mistake, it is a repeat (rule 4). If none applied, print nothing.
Hard rules
- Root cause is about Claude's process, not the artefact. "Off-by-one in the loop" is the symptom; "wrote the loop without testing the boundary" is the cause. Keep asking "what did I do or skip that let this through?" until the answer is an action Claude controls.
- The rule is checkable and scoped: an "Applies when" so it fires only on relevant work, and a "Check" answerable yes/no or by a command. "Be more careful", "double-check", "verify thoroughly" are banned.
- Log at the moment, never later. No apology paragraphs, no asking permission to log.
- Repeats escalate. If an exhibit's mistake happens again, the rule failed: increment Repeats and rewrite the rule and Check so they would have caught this case. At Repeats 2, propose moving the rule into CLAUDE.md, which is always loaded.
- No duplicates: same root cause means same exhibit, updated.
- Cap at 30 exhibits. When over, merge exhibits that share a root cause into one sharper rule, and move exhibits whose cause no longer exists to a one-line
## Retiredlist. - Never record secrets or personal data, and never record the user as the cause.
- Write "pass" only after actually running the Check.
First-time setup
When the file is first created, ask once whether to add this line to the project's CLAUDE.md, because that is what makes the museum load without being asked:
Read
.claude/failure-museum.mdbefore starting work and run its matching checks before finishing (skill: failure-museum).
Templates
File header:
# Failure Museum
Mistakes Claude has made in this project, each with a rule and a check that prevent a repeat.
Next ID: E01Exhibit (increment Next ID after adding):
## E07 · Edited generated dist/api.ts by hand; the next build wiped it
Date: 2026-03-02 · Repeats: 0
What happened: Hand-fixed dist/api.ts; the change vanished on the next build.
Root cause: Did not check whether the file was generated before editing it.
Rule: Before editing a file, check for a generated-file header or a build step that writes it.
Applies when: Editing anything under dist/, build/, gen/ or a file marked DO NOT EDIT.
Check: `git check-ignore <file>` and read the first 5 lines; edit the source instead if either shows it is generated.Non-code pair, to show the level required:
- Root cause: "Quoted a statistic from a search snippet without opening the source." Rule: "Open the cited page and find the figure before quoting it." Applies when: citing any number in writing.
Output format
Only these one-liners are printed; nothing else about the museum.
- Open:
Museum: matches E03, E07. - Admit:
Museum: added E07.(orMuseum: updated E07, Repeats 1.) - Check:
Museum check: E03 pass, E07 fail, fixed.
Anti-patterns
- Exhibits that describe the bug rather than the behaviour that allowed it.
- Vague rules with no Applies when or runnable Check.
- Reciting the whole museum at the start.
- Contrition essays.
- Logging preferences as mistakes.
- Batch-writing exhibits from memory at session end; that is
autopsy's job, and it writes its lessons here as exhibits. - Claiming a check passed without running it.
A stuck-detector loop caused by Claude's own error becomes an exhibit. For catching mistakes before they happen, use premortem or red-team; this skill only records and enforces what already went wrong.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

