# Learn

> Capture a lesson from a mistake into CLAUDE.md's Self-Learning Rules so the same mistake isn't repeated. Distills the most recent correction into ONE concise, reusable rule and appends it (deduped). Use right after Claude does something wrong and the user corrects it, or when the user runs /learn.

- Skill: `givepro91/learn` (Agent Skill)
- Install (CLI): `npx skillmds@latest add givepro91/learn`
- Raw SKILL.md: https://api.skillmd.com/api/skills/givepro91/learn/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: givepro91 (https://skillmd.com/u/givepro91)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/givepro91/learn

---


# /learn — append a rule from a mistake

Turn a correction into a durable rule in the canonical agent file's **Self-Learning Rules**, so the next session doesn't repeat it. This is the "compounding engineering" loop: every mistake makes the rules sharper.

## Procedure

1. **Find the canonical agent file** (the one holding the `CC-RULES` block). Project root = `git rev-parse --show-toplevel` (fallback: cwd). Resolve the same way `/claude-md` does: a `CLAUDE.md` symlink → its target; `CLAUDE.md` importing `@AGENTS.md` (or AGENTS.md canonical) → `AGENTS.md`; only `AGENTS.md` → `AGENTS.md`; else `CLAUDE.md`. In practice: pick whichever of `CLAUDE.md` / `AGENTS.md` actually contains `<!-- LEARN:ANCHOR -->`. If neither has it, tell the user to run `/claude-md` first (don't hand-edit).
2. **Determine the rule:**
   - If the user passed text after `/learn`, use that as the rule (lightly cleaned).
   - Otherwise, look at the **most recent correction** in this conversation — what Claude did wrong and what the user actually wanted — and distill it into **one** rule.
3. **Write the rule well** (this matters more than anything):
   - **Imperative and general**, not a story. ✅ "When editing a shared function, grep all call sites before changing its signature." ❌ "Earlier you broke the build."
   - **One line, one lesson.** Actionable next time. Prefer "always/never/before X do Y" shapes.
   - **Preserve the user's language** (Korean correction → Korean rule).
   - Capture the *generalizable* lesson, not the one-off detail.
4. **Append (deduped)** via the helper:
   ```sh
   node "${CLAUDE_PLUGIN_ROOT}/scripts/append-rule.mjs" <target file> "<the one-line rule>"
   ```
   (`<target file>` = the canonical file resolved in step 1.) It inserts `- (YYYY-MM-DD) <rule>` after the anchor (newest first) and skips near-duplicates.
5. **Confirm** the added line (or report it was a duplicate). Offer to commit (don't commit without approval).

## Examples

<good>
User: "you changed the API signature but didn't update the callers"
→ rule: "Before changing a function's signature, grep every call site and update them in the same change."
</good>

<good>
User (Korean): "스키마 바꾸기 전에 꼭 물어봐"
→ rule (Korean): "DB 스키마를 변경하기 전에 반드시 사용자 승인을 받는다."
</good>

<bad>
rule: "Don't make mistakes." → too vague, not actionable.
rule: "I broke the login test at 3pm." → a story, not a reusable rule.
</bad>

## Why a separate section + skill (not full automation)

Auto-detecting "mistakes" from chat is noisy and unreliable. `/learn` keeps a human in the loop (you decide a correction is worth a rule), while the helper makes the write deterministic and dedup-safe. The rules live in `CLAUDE.md` so they load every session and travel with the repo.