claudemd-lerner

name

claudemd-lerner

description

Install a per-repo learner that keeps the CLAUDE.md hierarchy and docs/ tree current from Claude Code sessions. Captures sessions into daily logs and, on a manual command or a 6-hour SessionStart gate, edits CLAUDE.md + docs/ in place (no knowledge wiki). Trigger when the user says "claudemd lerner", "install claudemd lerner", "keep CLAUDE.md current", "update my project docs automatically", "CLAUDE.md aktuell halten", "docs aktuell halten", or wants session-driven CLAUDE.md + docs/ maintenance that can seed an existing repo.

claudemd-lerner — Install Skill

Installs the claudemd-lerner engine into the current repo: SessionEnd/PreCompact hooks capture each session into <ldir>/daily/, and an updater applies those logs to the repo's CLAUDE.md hierarchy + docs/ tree (editing them in place). Updates run on a manual command or a 6-hour SessionStart gate. The SessionStart hook injects no context — CLAUDE.md and docs/ are already read at session start; it only fires the gated update.

The lerner dir (<ldir>/) holds ONLY machinery (hooks/scripts/_shared/daily). The actual outputs — CLAUDE.md files and docs/ — live at the repo root, the files the agent already reads. Nothing is ever written under .claude/. There is no knowledge wiki (that is the separate knowledge-compiler skill).

Authentication

The updater uses the Claude Agent SDK, which needs ANTHROPIC_API_KEY (or CLAUDE_CODE_OAUTH_TOKEN) in the environment. Subscription credentials are NOT sanctioned for third-party plugins — public/customer installs must set an API key. Capture (hooks/scaffold) works without it; only update/seed make API calls.

Naming / collision note

Invoke this as neurawork-cc-harness:claudemd-lerner. The name is collision-free, but use the fully-qualified form by convention. This skill coexists with neurawork-cc-harness:knowledge-compiler in one repo: its hooks use cl--prefixed filenames, so its .claude/settings.json entries never overwrite the compiler's.

Phase A — Recon (read-only)

Run the recon script and read its RECON_JSON blob:

python3 "${CLAUDE_PLUGIN_ROOT}/engines/claudemd-lerner/recon.py"

• status: NOT_A_GIT_REPO → stop; tell the user to run inside a git repo (offer git init).
• existing_ldir set → this is an ADOPT (refresh) install; reuse that name.
• Note suggested_depth, language, has_docs, seed_recommended, timezone, and clean.

Phase B — Ask

Use AskUserQuestion to confirm:

1. Lerner dir name — default claudemd-lerner (or the detected existing_ldir).
2. CLAUDE.md depth — default suggested_depth (1 = root CLAUDE.md only; 2 = root
   • immediate subdirs).
3. Docs dir — default docs.
4. Language — confirm the detected language (en/de).
5. Excluded dirs — confirm the suggested list (node_modules, .venv, dist, build, .git).
6. Timezone — confirm the detected one (display only; no hardcoding).
7. Seed now? — offer only when seed_recommended is true AND the tree is clean. Seeding builds/updates the initial CLAUDE.md + docs/ from the repo and makes API calls.

Write the chosen depth/docs/language/excluded into <ldir>/config.json after install (install seeds defaults; edit the file to match the answers).

Phase C — Execute

python3 "${CLAUDE_PLUGIN_ROOT}/engines/claudemd-lerner/install.py" \
  --lerner-dir <NAME> [--seed]
uv sync --directory <NAME>

Then tell the user to commit <NAME>/, .claude/settings.json, and any CLAUDE.md / docs changes the seed produced. After install:

• Sessions now capture automatically.
• Manual update: /neurawork-cc-harness:cl-update.