Aller au contenu principal
Autres

LESSONS Command

Lists the feedback memories captured for the current project (and globally) — the "lessons" learned from user corrections.

Context​

<arguments>

Objective​

Provide an overview of the self-improvement system: which rules, counter-examples, or preferences have already been recorded to avoid repeating the same mistakes.

Usage​

/lessons              # List all feedback memories for the project (read-only)
/lessons <keyword>    # Filter by keyword (e.g., /lessons test, /lessons git)
/lessons --promote    # Capture the lesson just learned into the personal store
/lessons --bootstrap  # One-off: backfill the personal store from existing memories
/lessons --prune      # Keep the personal store under its size budget

The --promote, --bootstrap, and --prune modes feed the personal cross-project lessons store at ~/.claude/rules/lessons.md (loaded into every project by Claude Code). They follow the self-improvement rule: every write is generalized + sanitized + human-confirmed, and the lessons are personal — never committed to any repo.

Modes​

--promote (explicit capture — fallback to the reflex)​

For when the in-conversation reflex didn't fire but you want to keep a lesson:

  1. Take the lesson just learned, generalize it to a one-line principle and sanitize it (strip project/company/person names, paths, URLs, identifiers, verbatim snippets, secrets).
  2. Check for recurrence — the store is in your context. If the same principle is already stored, propose bumping its (seen N times) marker instead of adding a duplicate ((seen 2 times) on the first repeat, then N+1).
  3. Show it and ask the user to keep / edit / discard (or confirm the bump).
  4. On confirmation, write to ~/.claude/rules/lessons.md (create if absent): append the new lesson as a - bullet under its matching ## Topic heading (create the heading if none fits), or apply the recurrence bump. If it can't be generalized, keep it as a local project memory instead.

--bootstrap (one-off backfill)​

Seed the store from lessons you already accumulated, so you don't start from scratch:

  1. Run claude-base lessons bootstrap-scan (lists existing feedback memories across all your projects as project<TAB>name<TAB>description candidates).
  2. For each candidate, decide if it is general/recurring (worth applying everywhere) or project-specific (skip). Generalize + sanitize the keepers.
  3. Propose the keepers and let the user confirm each; append the confirmed ones. This is a one-time action — not a background job.

--prune (stay within budget)​

  1. Run claude-base lessons prune-check (prints OK/OVER size/budget, DUP: lines for duplicates — section-aware, so ## Topic headings and a lesson's (seen N times) twin are handled — and RECUR N: for the most-repeated lessons).
  2. If OVER or duplicates exist, propose merges (near-duplicates, incl. folding a duplicate into a single (seen N times) line) and drops (superseded lessons). Surface the high-RECUR lessons as keepers. Never delete without explicit confirmation.

Workflow​

  1. Locate the memory directory
    • Per-project: ~/.claude/projects/<project-slug>/memory/
    • Global: ~/.claude/memory/ (if it exists)
  2. Filter feedback memories
    • Read MEMORY.md and each feedback_*.md file
    • Extract the type: feedback frontmatter for sorting
  3. Summary
    • For each memory, display:
      • Title + short description
      • Why (initial reason)
      • How to apply (when to apply the rule)
  4. Optional filtering
    • If a keyword is passed, keep only the memories whose title, description, or content matches

Expected output​

=== Feedback memories for this project ===

1. Manual review of infra PRs
   user reviews and merges infra PRs himself, no auto-merge even when CI is green
   Why: prior incidents with auto-merge missing context
   How to apply: never enable Dependabot auto-merge for infra repos

2. Claude Max works headless on user VMs
   `claude setup-token` is the official path for cron/CI on Max
   Why: Max plan supports headless via setup-token
   How to apply: do not default to Routines for automation projects

=== Global feedback memories (cross-project) ===
(none if ~/.claude/memory/ is empty)

Special cases​

SituationAction
No memoryInformative message + link to the auto-memory doc
Memory without type: feedbackIgnore (only list feedback)
Keyword filter with no matchMessage "No memory matches '<keyword>'"
Orphan memory (file without an entry in MEMORY.md)WARN, suggest re-indexing

IMPORTANT: This command is read-only. To add a lesson, the system prompt handles it automatically when the user makes a correction (signal detected by prompt-context.sh).

NEVER modify or delete a feedback memory without explicit confirmation from the user.

See also​