lessons-learned（吸取的教训）

Lessons Learned

Use this skill to maintain a lightweight project memory loop.

Treat this file as the single source of truth for lesson-memory behavior: trigger rules, recall flow, capture flow, limits, and validation. README.md, references/, and phase-entry commands must not override those rules.

Trigger Contract

Invoke this skill in these moments:

• Task start: Run recall before implementation.
• During task: Run capture when the user corrects your approach.
• Task end: Evaluate capture criteria and auto-capture qualifying lessons.

Do not trigger for:

• One-off factual Q&A with no reusable process.
• Pure concept explanation without actionable guidance.
• Formatting-only or administrative edits with no technical lesson.
• Non-reproducible context with no stable prevention rule.

Minimal Execution Guardrails

When this skill is the primary process skill in a session:

• For non-trivial work (3+ steps), create a task tracker before implementation.
• Before claiming completion, run relevant verification commands and report evidence.
• Do not replace verification with assumptions or partial checks.

Storage

Store all lesson artifacts under docs/lessons/ at the project root:

docs/lessons/
├── _index.md
├── api-timeout-retry-pattern.md
├── db-migration-run-order.md
└── null-check-before-save.md

Each lesson card is one atomic Zettelkasten note. Keep one reusable lesson per card.

Canonical Limits

Apply these limits everywhere in the skill package:

• Load 1-3 primary cards during recall.
• Load up to 2 related cards.
• Load at most 5 cards total.
• Use 0-2 related links per card.
• Use 3-6 tags per card.
• Keep confidence in the inclusive range 0.0-0.9.

If another file conflicts with these limits, this file wins.

Recall Phase

Run this phase before writing code.

1. Extract task keywords (technology, failure mode, domain terms).
2. Determine whether lesson storage exists:

- If docs/lessons/ does not exist, treat this as first-run state and skip recall. - If docs/lessons/ exists but _index.md does not, read existing card frontmatter directly for this recall pass and warn that _index.md should be rebuilt during the next capture/update maintenance pass.

1. Determine working scope from the task context:

- project for cross-cutting or repo-wide concerns - module for package/directory-level concerns - feature for a specific flow or component

1. Read docs/lessons/_index.md.
2. Build candidates for normal recall from cards in docs/lessons/.

- Exclude only cards that explicitly set confidence: 0.0. - Cards with missing confidence remain eligible and use legacy fallback. - Cards with confidence: 0.0 remain in docs/lessons/ for explicit historical lookup. - Load cards excluded from normal recall only when the user explicitly asks to review cards excluded from normal recall.

1. Rank eligible candidates with this order: tag match -> scope match -> confidence (desc) -> date (desc).

- Legacy fallback applies only when confidence is missing: derive it from source (user-correction=0.7, bug-fix=0.5, retrospective=0.3). If both are missing, use 0.3.

1. Load 1-3 primary cards.
2. Expand with related links from primary cards, loading up to 2 additional cards.
3. Enforce the hard cap: primary plus related cards must not exceed 5 total.
4. Apply loaded lessons as constraints for current work and mention loaded card IDs briefly.
5. Treat recall as read-only for confidence:

- Do not change confidence because a lesson was loaded. - Do not change confidence because a lesson was not loaded. - Do not change confidence because a lesson was not used recently.

If no cards match, continue work without lesson constraints.

Recall Warnings

Treat these as non-blocking and continue:

• docs/lessons/ is missing on first run.
• _index.md is missing, so this recall pass used direct card metadata.
• A loaded card is missing confidence and needs legacy fallback.
• A related target is missing.

Capture Phase

Run this phase when any of these conditions are met:

• The user corrected your approach
• The user asks for capture
• A bug fix revealed a reusable pattern
• A task completion surfaced a non-obvious reusable insight

Use this phase for both kinds of memory maintenance:

• creating a new lesson card when current work reveals a reusable new lesson
• updating an existing lesson card when current work shows that a recalled or related card is less applicable, partially outdated, or no longer valid

Step 1 — Decide whether to capture

Ask: *"Will this lesson save time next time a similar task appears?"*

Apply this step separately for two decisions:

• whether to create a new lesson card
• whether to update an existing lesson card

Skipping new-card creation does not skip corrective updates to an existing card. If current work shows that a recalled or related card needs correction, continue to Step 2 for that existing card even when no new lesson is captured.

Only capture non-obvious insights. Prioritize lessons such as:

• Hidden relationships between files/modules.
• Execution paths that differ from what the code structure suggests.
• Non-obvious config, env vars, flags, or ordering constraints.
• Breakthroughs where error messages were misleading.
• Tool/API quirks and required workarounds.
• Build/test commands or operational prerequisites that are not documented nearby.
• Files that must be changed together for correctness.

Capture if:

• ✅ Reusable pattern or decision rule
• ✅ Costly mistake with a clear prevention strategy
• ✅ Key parameter, config, or precondition that is easy to forget
• ✅ Multi-attempt solution (include failure reasons + success conditions)

Update an existing card if:

• ✅ current work shows a recalled or related card is less applicable than before
• ✅ current work shows a recalled or related card is partially outdated
• ✅ current work shows a recalled or related card should no longer participate in normal recall

Skip if:

• ❌ One-off Q&A with no reusable process
• ❌ Pure concept explanation without actionable guidance
• ❌ Non-reproducible, context-specific conclusion
• ❌ Obvious framework/language behavior with no hidden constraint
• ❌ Session-specific notes that are unlikely to repeat

These skip rules apply to creating a new card. They do not block corrective updates to an existing card when current evidence shows that the card should be rewritten or its confidence should change.

Step 2 — Write the Zettel card

Generate a semantic kebab-case ID that describes the lesson (e.g., api-timeout-retry-pattern) when creating a new card.

Before writing or updating a card, check for a semantically similar existing card and make the decision explicit:

• decision=create when no similar card exists and a new lesson should be captured.
• decision=update when a similar card already covers the same lesson, or when current work corrects an existing card.

Use decision=create or decision=update only when all affected cards in the capture pass share the same outcome kind. If one capture pass both creates and updates cards, report only the created=<n> and updated=<n> counts.

When one capture pass affects multiple cards, report created=<n> and updated=<n> counts. If one capture pass both creates and updates cards, do not include decision=.

Assign a scope before writing the card:

• project for repo-wide constraints
• module for package/directory-level constraints
• feature for one flow/component

Assign initial confidence by source:

• user-correction: 0.7
• bug-fix: 0.5
• retrospective: 0.3

Confidence changes only during capture/update, never during recall.

If the user explicitly confirms a lesson remains useful, increase confidence by +0.1 (max 0.9).

Decrease confidence by -0.1 when later evidence shows the lesson is less applicable.

When a lesson is partially outdated, update the card content first, then decrease confidence only if needed.

Set confidence directly to 0.0 only when the card should no longer participate in normal recall.

Do not decrease confidence solely because the lesson was not used recently.

Write the card to docs/lessons/<id>.md using the template in references/card-template.md.

Before creating a new card, check semantic duplication:

• If an existing card is semantically similar, update that card instead of creating a duplicate.
• Preserve existing card ID when updating.

When current work weakens a recalled or related card:

• update the existing card even if no new lesson card is created
• update the card content first when the lesson is only partially outdated
• then lower confidence only if the lesson's default applicability is weaker than before
• set confidence to 0.0 only when the card should no longer participate in normal recall

Minimal correction-capture example:

Lessons capture report: decision=update, updated=1 (db-migration-run-order), confidence=0.7->0.6 (weaker applicability in current config), skipped=0

Card fields:

Task-end behavior:

• If capture criteria are met, auto-capture without asking for permission first.
• After capture, report what was created or updated.
• For normal user-facing output, prefer a short capture summary.
• Do not reproduce the full card body unless the user asks to inspect it.

Step 3 — Update the index

Upsert one row in docs/lessons/_index.md per card ID.

If _index.md does not exist, create it with this structure:

# Lessons Index

> Auto-generated by lessons-learned skill. Do not edit manually.

| Card | Scope | Tags | Date |
|---|---|---|---|

Then keep rows sorted by Date descending (newest first).

Step 4 — Confirm with user

Tell the user what was captured using a compact report, e.g.:

Lessons capture report: created=1 (api-timeout-retry-pattern), updated=1 (db-migration-run-order), skipped=1 (obvious behavior)

If all affected cards share the same outcome kind, include decision=create or decision=update in the report. If one capture pass both creates and updates cards, report only the created=<n> and updated=<n> counts.

When capture or update changes confidence, include confidence transition old->new for each affected card plus a short reason.

For new cards, use none-><value>.

If a transition reaches 0.0, explicitly state that the card is excluded from normal recall.

Examples:

• Lessons capture report: decision=create, created=1 (\api-timeout-retry-pattern), confidence=none->0.5 (new reusable pattern)
• Lessons capture report: decision=update, updated=1 (\db-migration-run-order), confidence=0.7->0.6 (weaker applicability in current config)
• Lessons capture report: decision=update, updated=1 (\legacy-bootstrap-flow), confidence=0.1->0.0 (excluded from normal recall)

Keep this confirmation short. Prefer the decision, the affected card ID, and a one-line rule summary. Avoid dumping the full markdown card or index contents in normal output.

Linking Rule

Use related links for high-value cards only.

Create related links when at least 2 of 4 conditions are true:

1. Reusable across different tasks.
2. Represents a high-cost mistake (multiple attempts or significant time loss).
3. Depends on critical parameter/config/decision details.
4. Connects naturally to at least two existing lesson cards.

Linking constraints:

• Add 0–2 related links per card, only when relevance is strong.
• Use deterministic ranking from references/linking-heuristics.md.
• Do not add weak or speculative links.

Broken related targets are non-blocking:

• If a related target is missing, ignore it and warn.
• Continue recall/capture flow.

Anti-patterns

• Capturing obvious language/framework behavior with no hidden constraint.
• Capturing session-only notes (temporary logs, ad-hoc local paths, one-off environment noise).
• Writing verbose narrative cards without actionable prevention rules.
• Creating near-duplicate cards when an update to an existing card is enough.
• Recording broad advice without a concrete trigger in When to Apply.

Validation

Apply these checks during capture or update. These are blocking failures unless explicitly listed as warnings elsewhere.

• Card filename equals id slug.
• date format is ISO YYYY-MM-DD.
• scope is one of project, module, feature.
• tags count is 3–6.
• source is valid enum.
• confidence is numeric and in range 0.0-0.9.
• related count is 0–2 and every target resolves to an existing card.
• Index row exists and matches card metadata, including scope.
• _index.md rows are ordered by Date descending (newest first).

Integration Guide

When used with other skills in the same session, follow the Trigger Contract as the single source of truth:

• task start -> recall
• user correction during work -> capture
• task end -> capture evaluation

This is a non-replaceable step — lesson capture cannot be substituted by todo trackers, progress files, or other skill artifacts.