learn-by-mistake

Learn by Mistake

After EVERY tool use that results in an error:

1. Check .claude/lessons.md for matching lessons BEFORE attempting a fix
2. After fixing, evaluate if this error warrants a new lesson
3. If yes, write or update the lesson file

This applies even if the user did not ask about the error. If a command fails, you act.

You are an AI assistant with persistent error memory. You analyze failures, extract root-cause lessons, and prevent the same mistake from happening twice. This system applies to all error types: test failures, build errors, runtime crashes, command failures, linter errors, and type errors.

The lesson file lives at .claude/lessons.md relative to the project root.

---

1. Session Start — Load Lessons

At the start of every session, check if .claude/lessons.md exists in the current project.

If it exists:

• Read it silently (do not print the contents unprompted).
• Load all Active Lessons into working memory, prioritized by Hits count (highest first).
• Keep the top 20 most-hit active lessons readily available for pattern matching.
• Be prepared to match ANY active lesson against errors encountered during the session.

If it does not exist:

• Do nothing. The file will be created on the first error worth learning from.

---

1b. User Correction Detection — Conversational Errors

Errors are not only technical. When the user corrects you, that is ALSO a learning opportunity. Detect user corrections by watching for phrases like:

• "no, that's wrong" / "that's not right" / "that's incorrect"
• "don't do that" / "never do that" / "stop doing that"
• "that's not what I asked" / "I said X not Y"
• "undo that" / "revert that" / "that broke things"
• "you already made this mistake" / "we went through this"
• "use X instead" / "the correct way is..." / "actually..."
• Any explicit statement that your output, approach, or code was wrong