| name | retro-notes |
| description | Capture AI-readable postmortems after a dev attempt resolves into a project-scoped lessons file that future /big-plan runs can read. Use when: (1) User says 'retro-notes', 'retro', 'capture lessons', 'fail-notes', 'postmortem', (2) A dev session wrapped up that took longer than expected, changed approach mid-way, or had non-obvious gotchas. Reader is a future planning agent. |
Retro Notes
Purpose
Convert lessons from a just-resolved dev attempt into a structured, project-scoped lessons file (/l-lessons-{area} skill) that future planning runs (e.g. /big-plan) can consume. Goal: turn one-time pain into permanent leverage so the next attempt in the same area is faster.
Pre-flight check (mandatory)
Before writing, confirm two things:
- Is the work actually wrapped up? If the user is still mid-struggle, ask them to confirm. Mid-struggle notes record symptoms, not root causes — they crystallize the wrong lesson. If unresolved, suggest they invoke this skill again after the fix lands.
- What's the project area? Look at the current working directory, recently touched files, and conversation topic. Propose a project-area name (kebab-case, e.g.
canvas-tools, tauri-window-mgmt, react-form-state). Confirm with the user before writing.
Locate or create the lessons file
The output is a project-scope skill at .claude/skills/l-lessons-{area}/SKILL.md.
One growing file per problem-area, NOT one file per attempt.
- Check
.claude/skills/ for any existing l-lessons-* skill that matches the area. If a relevant one exists, append a new dated section to its SKILL.md.
- If none exists, scaffold a new skill at
.claude/skills/l-lessons-{area}/SKILL.md with the frontmatter shown below.
Frontmatter for new lessons files
---
name: l-lessons-{area}
description: Project lessons learned for {area}. Read PROACTIVELY before planning or implementing work touching {area} — contains traps, root causes, and "watch for next time" notes from previous attempts.
---
Keep description short but concrete enough that planning agents pick it up automatically.
Note template
Append (or create) a section using exactly this structure:
## {YYYY-MM-DD} — {short topic}
### What we set out to do
{1–2 sentences. The original goal, not the implementation.}
### Approach we tried first
{The wrong path. Be specific about the abstraction or structural choice — not the symptom.}
### Why it went wrong (root cause)
{The structural cause, not the symptom. See "Symptom vs root cause" below.}
### What worked instead
{The right abstraction or approach.}
### Watch for next time
- {Concrete trap, ideally in "if you see X, you're probably Y" form.}
- {Another trap. Aim for 2–5 bullets, not narrative.}
### Would-skip-if-redoing
{Things that wasted time and are now provably unnecessary. One short paragraph or bullet list.}
The Watch for next time section is the highest-leverage part. Future planning agents will scan it first. Keep bullets concrete and trap-shaped.
Symptom vs root cause
When the user offers a symptom-shaped description, probe for the structural cause before writing.
| Symptom (what the user says first) | Root cause (what to actually record) |
|---|
| "zoom was buggy" | "transform threaded through every function instead of inverted at input boundary" |
| "form kept losing state" | "state owned by child instead of lifted to parent that survives re-mount" |
| "the build was flaky" | "test relied on filesystem ordering not guaranteed by the runtime" |
| "performance was bad" | "N+1 query in the render loop, hidden behind an inner abstraction" |
If the user gives a symptom, ask: "What was the actual structural reason that made the symptom show up — what was wrong about the shape of the code, not the bug itself?" Record the structural answer. The symptom can stay as one phrase in What we set out to do or Approach we tried first for context.
Writing style
- Terse. The reader is a future LLM scanning for relevant patterns, not a human reading prose. Cut adjectives, cut narrative.
- Specific over general. "Coordinate transforms must invert at input boundary" beats "be careful with coordinates."
- Past tense, no apology. "We threaded the transform through every function" — not "I made a mistake by..."
- No grep-able blame. Don't name people or PRs; the lesson is structural.
After writing
Briefly tell the user:
- The lessons file is at
.claude/skills/l-lessons-{area}/SKILL.md. /big-plan auto-reads matching l-lessons-* skills in its Step 1c before planning, so future planning runs in this area pick it up automatically.
- Suggest a periodic
/retro-notes-refactor pass (e.g. once the lessons file accumulates 5+ entries): merge overlapping notes, delete advice for code that no longer exists, promote repeated lessons into broader wisdom.
- If the retro produced "nothing surprising, plan matched reality," that's signal too — the area is now well-mapped, future planning in this area can be lighter.
When NOT to invoke this skill
- The work is still in progress or stuck.
- The dev was trivial and went exactly as planned (no lesson to extract).
- The lesson is universal CSS/JS knowledge that belongs in a global wisdom skill (e.g.
/css-wisdom), not a project-scoped one.
