← ClaudeAtlas

learninglisted

Turns a repeated mistake or correction into the smallest lasting fix (CLAUDE.md line, rule, hook) and prunes stale rules. Use when the user says you keep doing X, wants a rule for this project, or asks to clean up CLAUDE.md. Proposes first.
t8jwscc7qy-sketch/foma-skills · ★ 2 · AI & Automation · score 78
Install: claude install-skill t8jwscc7qy-sketch/foma-skills
Finds what this session taught about the project, puts each lesson in the smallest layer that fixes it, and removes rules that no longer earn their place. Propose, then write after the user's ok. ## Steps 1. *Always.* Collect evidence from this session and `$ARGUMENTS`: a user correction, the same failure twice, a task that took several attempts, a repeated workflow. No concrete event and no request to prune: say "nothing to learn" and stop. 2. Read what exists: the CLAUDE.md contract and rules, `.claude/rules/`, hooks in `.claude/settings*.json`, project skills in `.claude/skills/` (names and descriptions only), the decisions file. `grep` for the lesson's keywords first: one home per fact. 3. Gate each lesson; drop it if any answer is no: a concrete event backs it; the agent cannot infer it from the code; no rule or hook already covers it. 4. Pick the smallest layer that works: nothing (one-off) → a CLAUDE.md line (one sentence, "Claude gets this wrong: …") → a path rule in `.claude/rules/` → for a rule that must hold with no exceptions, a `permissions.deny` entry ("never touch this path") or a hook (Bash hooks block, Write/Edit hooks only warn) → a project skill (a recurring multi-step procedure only). A change bigger than that: add a `proposed` entry to the decisions file and name `/foma-skills:groundwork` or foma-skills:shaping; do not do it here. 5. Prune in the same pass: rules naming files, commands or tools that no longer exist (check with grep), duplicates, contradi