Lesson refresh

Lesson Refresh

Core Rule

Decide keep / update / promote / encode / archive per lesson based on relevance and recency. Never delete a lesson without recording why.

Kit Context

Before starting this skill, ensure you have completed session boot:

1. Read CODEBASE_MAP.md for project understanding
2. Read CLAUDE.project.md if it exists for project-specific rules
3. Read tasks/lessons/_index.md for accumulated corrections (Top Rules + index)

If any of these haven't been read in this session, read them now before proceeding.

When to Use

Invoke with /lesson-refresh when:

• It's been ≥2 weeks since the last refresh (or never)
• The Top Rules index has grown beyond ~10 items
• You've shipped a major refactor that may have invalidated old lessons
• The session-boot context feels noisy (compaction kicking in too often)
• You want to promote a recurring lesson into a hook or CLAUDE.md rule

Not for:

• Adding new lessons — those come from user corrections during regular sessions
• Editing a single lesson — just open the file
• Project-wide cleanup — that's the broader "Spa Day" ritual in agent_docs/skills.md

Scope Rules

• Operates only on files in tasks/lessons/
• Reads CLAUDE.md, CLAUDE.project.md, .claude/hooks/, and code surface as evidence — never modifies them
• Modifies lesson frontmatter (status, updated, top_rule) and _index.md Top Rules
• Moves archived lessons to tasks/lessons/_archive/ (creates if missing)
• Never deletes lesson files — only sets status: archived and moves them

Process

Phase 1: Inventory

1. List all tasks/lessons/*.md files except _index.md and _TEMPLATE.md
2. For each, read frontmatter and body
3. Group by status: active, archived, superseded
4. Skip non-active lessons unless --include-archived is in arguments (rare; for audits)

Phase 1.5: Graph signals (typed-relation read)

Run ./scripts/lesson-graph.sh --check and read the warnings it emits. Each warning is a deterministic signal the refresh should react to before consulting any other evidence:

Graph warning	Verdict bias
supersedes-target-missing: A supersedes X (missing)	The pointer is stale — open A, remove X from supersedes: or correct the slug. Apply as update.
contradicts-target-missing: A contradicts X (missing)	Same — clean up the pointer. Apply as update.
supersedes-cycle: A → B → A	Two lessons claim to replace each other. Reconcile manually: decide which one is the survivor, set the loser's status: superseded, drop its supersedes: clause.
contradicts-loop: A ↔ B	Two lessons disagree. Surface to user — the resolution is content, not graph: either one is wrong, or the rule needs sharpening.
top-rule-but-superseded: A	Strong archive or update: the lesson is both promoted and outdated. Demote top_rule: false and pick archive or keep based on its remaining utility.

After Phase 1.5, regenerate the index unconditionally — even if no verdicts apply — so the Active Rules By Topic / Recently Added / Superseded sections stay fresh:

./scripts/lesson-graph.sh

Phase 2: Evidence Gathering

For each active lesson, gather signals to inform the verdict:

Signal	How to check	What it means
Code surface unchanged	Grep for the file/function the lesson names; does it still exist?	If the path is gone, the lesson may be moot
Encoded as hook	Scan .claude/hooks/ for a rule matching the lesson's Rule line	Lesson may be redundant — agent can't violate the rule anymore
Encoded in CLAUDE.md	Search CLAUDE.md and CLAUDE.project.md for the rule	Lesson is encoded as a top-level rule; duplication
Encoded in skill	Search .claude/skills/*/SKILL.md for the same rule	Lesson covered by a discoverable skill
Recurrence	Search tasks/lessons/ for sibling lessons with overlapping tags or text	Indicates a pattern worth promoting
Age	created field vs today	Stale lessons (>180 days, untouched) need a closer look
Update count	If updated > created by a wide margin	Lesson is still actively relevant

Capture the signal set per lesson — do not act yet.

Phase 3: Verdict

For each lesson, decide one of:

Verdict	When	Action
keep	Still relevant, not encoded elsewhere, no signal that it's stale	No frontmatter change. Log in report.
update	Still relevant but message/rule needs sharpening, tags are wrong, or scope drifted	Edit the file's body and/or tags; set updated: today
promote	Recurring pattern (2+ sibling lessons), should appear in Top Rules	Set top_rule: true; add a one-line entry to _index.md → ## Top Rules; set updated: today
encode	Rule belongs in CLAUDE.md or a hook, not in the lesson archive	Propose the encoding (don't apply it without user approval). Set status: superseded. Add a superseded_by: <path> field in frontmatter pointing to the encoded location once approved.
archive	Code/dep gone, lesson moot, or duplicated elsewhere	Set status: archived; move the file to tasks/lessons/_archive/; remove from _index.md Top Rules if listed

A lesson can match only one verdict per refresh. If two verdicts seem plausible (e.g., promote vs encode), prefer the one with the lower long-term token cost (encode > promote > keep > update > archive in terms of effort, but encode is best because it eliminates the lesson from session boot context).

Phase 4: Apply (Interactive Mode)

Present the verdict batch to the user with a compact table:

# Verdicts (N lessons reviewed)

| Lesson | Verdict | Why |
|---|---|---|
| 2026-04-30-tailwind-purge-cache.md | keep | Rule still applies; no encoding found |
| 2026-04-15-tsconfig-wrong-edit.md | encode | Pattern hookable; suggest .claude/hooks/protect-files.sh enhancement |
| 2026-03-10-async-state-trap.md | archive | Code path removed in commit abc1234 |
| 2026-02-20-cors-config.md | promote | Recurrence: 3 sibling lessons in tag "auth" |

Apply all? (y/n/select)

• y — apply every verdict
• n — abort, no changes
• select — drop into per-lesson confirmation; user can override each

After applying:

• Print a summary: counts per verdict, paths touched
• For encode verdicts, write the proposed encoding (hook diff or CLAUDE.md addition) into tasks/todo.md > ## Lessons to Encode so the user can review without losing context

Phase 5: Apply (Headless Mode)

In headless mode, apply only safe verdicts:

• keep — no change
• update — apply only if confidence in the new tags/scope is high
• archive — apply only if the evidence is "code surface gone" (file path doesn't exist anymore)

Skip in headless mode:

• promote — promoting to Top Rules affects session boot for everyone; should be human-approved
• encode — proposes changes to CLAUDE.md or hooks; never silent

Save the verdicts that were skipped to tasks/todo.md > ## Lesson Refresh Pending so the user can review on next interactive session.

Run Mode

This skill supports interactive (default) and headless modes — see the canonical contract in .claude/skills/_shared/blocks/mode-detection.md.

Headless detection: presence of mode:headless in arguments.

Decision point	Interactive default	Headless default
Apply all vs per-lesson (Phase 4)	Show batch table, ask y/n/select	Apply only safe verdicts (Phase 5); queue the rest to tasks/todo.md > ## Lesson Refresh Pending
Promote / encode	Suggest, await approval	Skip — queue for human review
Archive ambiguity (lesson might still be relevant but old)	Ask	Keep, don't archive — bias toward retention in headless mode
End	Summary + "anything to revisit?"	Print summary table only

Output Format

# Lesson Refresh Report — YYYY-MM-DD

**Reviewed:** N active lessons (M archived skipped)
**Verdicts applied:** K keep, U update, A archive, P promote, E encode

## Verdicts

| Lesson | Verdict | Action | Why |
|---|---|---|---|
| 2026-04-30-tailwind-purge-cache.md | keep | — | Still applies; no encoding found |
| 2026-04-15-tsconfig-wrong-edit.md | encode | queued in tasks/todo.md | Hookable; better as PreToolUse check |
| 2026-03-10-async-state-trap.md | archive | moved to _archive/ | Code path removed in commit abc1234 |

## Pending (Headless mode only)

These verdicts require human review — queued in `tasks/todo.md > ## Lesson Refresh Pending`:

- promote 2026-02-20-cors-config.md (recurrence: 3 sibling lessons)
- encode 2026-04-15-tsconfig-wrong-edit.md (hook candidate)

## Top Rules Diff

Before: N rules
After: M rules
Added: <list>
Removed: <list>

Notes

• This skill is the counterpart to the kit's broader "Spa Day" cleanup ritual (see agent_docs/skills.md). Spa Day reviews CLAUDE.md, agent_docs, and skills; lesson-refresh focuses only on tasks/lessons/.
• The encode verdict never modifies CLAUDE.md or hooks silently — it always queues for human approval. This is a load-bearing invariant: encoded rules affect every session and must be human-vetted.
• The _archive/ directory exists so archived lessons are still findable. They are excluded from _index.md and from on-demand reads, but discoverable via grep when investigating past incidents.
• Recommended cadence: run interactively every 2–4 weeks, or schedule a monthly headless run via /loop for the safe-verdict pass.