Learn

Learn from Corrections

Review the current conversation for corrections, extract generalizable rules, and persist them.

Workflow

1. Identify Corrections

Look for:

• Explicit corrections: "No, don't do X" / "That's wrong, it should be Y" - extract the rule and its positive alternative
• Implicit corrections: User rewrites output, ignores a suggestion, or asks again with more specificity
• Positive confirmations: "Yes, exactly like that" / acceptance of a non-obvious approach

2. Extract the Principle

For each correction:

• Is it generalizable? "Use bullet points not paragraphs" is. "Change this sentence" is not.
• What's the scope? All agent behavior, a specific skill, a specific tool?
• Is there a "why"? Rules with reasons are more useful than bare rules.

Skip one-time, task-specific corrections.

3. Route to the Right File

Type	Target
Global agent behavior	CLAUDE.md
User preferences	memory/feedback_*.md
Skill-specific	.claude/skills/*/SKILL.md
Domain knowledge	{domain}/knowledge.md

4. Implement the Change

• Memory files: Follow the frontmatter format (name, description, type: feedback). Structure as: rule, then Why: and How to apply: lines. Update memory/MEMORY.md index.
• CLAUDE.md: Add to the relevant section, one line per rule.
• Skills: Add to the skill's Key Principles or update the workflow step.

Check existing files first - update rather than duplicate.

5. Verify and Summarize

Read back every modified file. Check frontmatter integrity and markdown validity.

Present a summary for each correction:

## Learned

### [Correction]
- **What was corrected:** [wrong behavior]
- **Principle:** [the rule]
- **Why:** [user's reasoning]
- **Updated:** [file path]

Key Principles

• Rules need reasons. Always capture the why.
• Update, don't duplicate. Check existing files first.
• Positive patterns matter. Codify confirmed approaches, not just mistakes.
• Scope tightly. Skill-specific rules stay in skills, not CLAUDE.md.