Persist corrections and patterns to long-term memory. When a mistake is made or a rule is learned, write it to .claude/state/learnings.md so it's never repeated.
Install
npx skillscat add strikersam/autonomous-ai-agency/learn-rule Install via the SkillsCat registry.
SKILL.md
Skill: learn-rule
When to Use
Use this skill whenever:
- The user corrects a mistake you made
- You discover a repo-specific rule that isn't in CLAUDE.md
- A pattern works unexpectedly (positively or negatively)
- You want to persist a decision for future sessions
This is how Claude gets smarter at this specific repo over time.
Instructions
Step 1 — Identify the rule
Extract the core rule from the correction or discovery. A good rule is:
- Specific — refers to a concrete situation
- Actionable — tells you what to do or not do
- Memorable — can be read in under 10 seconds
Bad rule: "Be careful with auth code."
Good rule: "Always call risky-module-review before touching admin_auth.py — even for 1-line changes."
Step 2 — Append to learnings file
File: .claude/state/learnings.md
Create it if it doesn't exist. Always append — never overwrite.
## YYYY-MM-DD — <short title>
**Situation:** What was being worked on when this was learned.
**Mistake/Discovery:** What went wrong or what was surprising.
**Rule:** <One clear, actionable sentence starting with a verb.>
**Source:** user-correction | self-discovery | docsExample entry:
## 2026-04-09 — Don't use git add -A
**Situation:** Committing the router refactor.
**Mistake:** Used `git add -A` which staged keys.json accidentally.
**Rule:** Always stage specific files by name; never use `git add -A` or `git add .`.
**Source:** user-correctionStep 3 — Check if CLAUDE.md should be updated
If the rule is general enough to apply to every future session:
- Read CLAUDE.md
- Add it to the "Coding Rules" section (numbered, concise)
- Use
repo-memory-updaterskill if multiple rules need updating
Learnings File Format
.claude/state/learnings.md:
# Session Learnings
Rules and corrections learned from working in this repo.
Read this at the start of sessions via the `replay-learnings` skill.
---
## YYYY-MM-DD — title
**Situation:** ...
**Mistake/Discovery:** ...
**Rule:** ...
**Source:** ...
---Acceptance Checks
- Rule is specific and actionable (not vague advice)
- Entry appended to
.claude/state/learnings.md - CLAUDE.md updated if rule is broadly applicable
- Rule phrased as an imperative: "Always X", "Never Y", "Before Z, do W"