Tandem: reflect (self-improvement)

Use at the end of a substantive session (>30 min of real work) to capture lessons that should bleed back into the kit's rules. The blog's "self-improvement loop" pattern.

Ground the scan in the ledger first

Before scanning from memory, read what the run recorded at the time. The retro ledger holds one line per development boundary — the friction, the kit signals and the estimate-vs-actual calls each close-out wrote as it happened — and recalling those is stronger evidence than reconstructing them at the end of a long session:

node _00-Project-Management/93-Scripts/retro-report.js --month <YYYY-MM>

Read the current month only — that is the in-flight view this skill needs, and the ledger grows monotonically, so loading all of it would cost more than it is worth here. This is 93-Scripts/retro-report.js, the same aggregator close-phase and monthly-retro read: reflect does not re-implement the join, so the three surfaces cannot disagree about what happened.

An empty window changes nothing about this skill. The output says No retro records for this window.; carry on with the scan below exactly as before. No error, no banner, and nothing attributed to a ledger that has nothing in it — a reflection with no ledger line to cite is still a valid reflection, and every session before 2026-08 has none.

What to scan in this session

1. Decisions made — did you create an ADR? Did you make a decision that should have become an ADR but didn't? (the rule says "any non-obvious decision" — be honest.)
2. Friction encountered — did you have to explain the same convention to Claude twice? That's a PROJECT-CONTEXT.md candidate.
3. Workarounds applied — did you bypass a rule or use a one-off command? Why? Should the rule change, or should there be a new entry in PROJECT-CONTEXT.md's "Known stack gotchas"?
4. Skills that didn't fire — did Claude paste a prompt's content instead of loading a skill? The skill's description: may not match the trigger phrase the user actually used.
5. Tools used that surprised you — was an agent invoked when a direct Read would have been faster? Was main-thread context bloated by grep results? Subagent policy adjustment needed?
6. Patterns that recurred — did you tell Claude to do something three times that could be a hook?

What to output

A proposal — not a commit. The user decides what lands.

🪞 Session reflection — <ISO date>

What worked:
  - <specific thing, ≤2 lines>

What hurt:
  - <specific thing, ≤2 lines>

Proposed kit changes (review before applying):
  1. [PROJECT-CONTEXT.md] Add to "Known stack gotchas": <symptom> — <fix>. Reason: <one line>.
  2. [SOP.md §<N>] Tighten rule: <current text> → <proposed text>. Reason: <one line>.
  3. [skills/<name>/SKILL.md] Update description to include trigger phrase: "<phrase user actually used>". Reason: <one line>.
  4. [hooks/hooks.json] New hook candidate: <event> running <command>. Reason: <one line>.

ADR backlog (decisions made this session without an ADR — file these now):
  - <decision> made at <timestamp/commit>. Should be ADR-<NNNN>.

Tech debt observed (file as BACKLOG entries):
  - <observation>

User confirmation needed to apply any of the above.

Output rules

• Propose, don't commit. No file edits during reflect. The user reviews and explicitly approves each item before it lands.
• One proposal per finding — not bundled.
• Cite evidence: "I noticed X at <approx point in session>" — concrete, not vague.
• Skip the section entirely if there's no finding for that category. Don't pad.
• ≤ 30 lines total. If the session yielded more than 5 findings, prioritise the top 3 and note the rest as "additional minor findings: <count>".

Anti-patterns

• Proposing rule changes for one-off situations. The bar is "this would have helped me twice or more in the past month."
• Suggesting new skills when a richer description on an existing skill would solve it.
• Auto-applying proposals without user approval — even small ones. The user owns the kit.

Non-negotiable rules from CLAUDE.md

• Read-only — reflect proposes; user applies.
• Subagent delegation for multi-file scans of past sessions.
• If a proposal touches _00-Project-Management/90-Standards/SOP.md, also bump the file's version: field in frontmatter once the user approves.

End-of-reflect summary

A single line: "X proposals · Y ADRs to backfill · Z tech-debt items. Apply now? (y/n)"