learn-from-mistake

learn-from-mistake

After being corrected, proactively propose solidifying the lesson, then generate a guardrail rule upon user approval.

Workflow

1. Identify the error: Triggered when the user corrects the AI. Extract the following:

   • Error behavior: what the AI did
   • Trigger condition: under what circumstances it happened
   • User correction: what the user expects as the correct behavior
   • If the user's correction is unclear, confirm first: "I understand you think [X] is the problem, and the correct approach should be [Y] — is that right?"

2. Summarize and propose: Abstract the specific error into a general pattern, present it to the user, and ask for their opinion:

   • Use the pattern: "When X, the AI tends to Y, but the correct approach is Z"
   • Show the summary directly so the user can decide whether to solidify based on the concrete content
   • Stop if the user disagrees

3. Generalization self-check: After summarizing the pattern, run a self-check to ensure appropriate granularity:

   • X describes an identifiable coding/operational scenario, not something vague like "when writing code"
   • Z describes an executable check step, not "be more careful"
   • If the self-check fails, adjust the granularity — err on the side of being too specific rather than too abstract

4. Determine output form:

   • Requires an "incorrect vs correct" behavior/operation/code comparison to explain clearly → generate a skill
   • A rule that can be stated in one sentence → update CLAUDE.md
   • When both are needed, do both

5. Generate output: Execute according to the determined form

6. Quality assurance: If the output is a skill, call skill-creator:skill-creator before writing the file to evaluate the description's trigger coverage and content quality, then revise based on the evaluation before writing

Generate a Skill

Place it in .claude/skills/{name}/SKILL.md in the current project root (not the user home directory ~/.claude/skills/).

Before generating, check whether a similar skill already exists under .claude/skills/ — update it rather than creating a new one.

Content Guidelines

The generated skill should include the following core elements (format is flexible; organize based on the nature of the error):

• Trigger condition: under what circumstances this mistake tends to occur
• Error behavior: what the AI typically does (can be illustrated with code, operational procedures, behavioral descriptions, etc.)
• Correct approach: the corrected approach (also not limited to code)
• Check steps: 2–4 actionable check actions that the AI can actually perform during coding

Requirements

• Skill name in kebab-case
• Error patterns should be generalized, not limited to the current case only
• Correct approach should be specific enough for the AI to follow directly
• Check steps should be actionable — each step is a concrete action

Update CLAUDE.md

Append to the project's CLAUDE.md. If CLAUDE.md does not exist, create it first with a basic structure.

When appending, find the semantically most relevant section to insert into — do not pile everything at the end of the file. If no suitable section exists, create a new one.