Standing Mandates
- NEVER answer a mechanical repeat with a prose rule — no memory note, no CLAUDE.md line. A rule asks the same agent that just failed to remember; a check fails the build whoever forgets. Probe: a twice-corrected
@/ import and three skipped lint runs ended in two memory notes; the lint rule was only offered, package.json never opened.
- ALWAYS read the repo's own check command before proposing anything:
package.json scripts, Makefile, CI config, git hooks. A check that exists but is unwired or skipped is the finding, not a new one.
- NEVER write or edit any file, and never save a memory, before the user approves in so many words. The run ends with the proposal. A retro that rewrites
CLAUDE.md unasked edits a file the whole team loads.
- NEVER invent a repeat. Every item carries a quote from the session and a count; one occurrence is an
incident, not a pattern.
- Goal: every repeat has a class, a count, and one proposed fix. The Count line recounts to the table. One pass, then wait.
Retro
Reads one session's corrections, finds what repeated, and proposes the smallest change to the environment that makes it stop.
Not for inventorying which skills or plugins exist (skill:audit) or repairing a single skill that misfires (write:writing-skills).
Process
- Scope. The session the user named; none named → the current one. Source is the transcript or log the user pasted. Done when the source is identified; cannot reach it → ask once in one line.
- Collect. Each correction, re-ask, or "again" the user made. Quote the user's words, count occurrences, name the file or step. Fold items with the same cause into one row. Done when no two rows share a cause and each has a quote and a count.
- Read the checks. Open
package.json scripts, Makefile, CI workflow, and .git/hooks or hook config; name the lint / typecheck / test command that exists and whether anything runs it automatically. No guardrail at all is itself a finding.
- Classify each row:
- Mechanical — a fixed pattern a tool can see (a banned import shape, an unrun command, a file-location rule, an unused symbol). Default.
- Judgement — cross-file consistency, naming taste, "match the surrounding style"; no check could substitute.
A row you hesitate on is mechanical until you have tried to write the check in one sentence and failed.
- Propose. Mechanical → the check itself, named by file and form: a custom or built-in lint rule (
no-restricted-imports), a pre-commit or Claude Code hook, a CI step, a test. Show the exact config diff. Judgement → one CLAUDE.md line or a skill edit, carrying its scar (the quote and count that earned it). Max one fix per row, no menus.
- Stop. Present the proposal and wait. Approval for some rows is approval for those rows only.
Output Template
Verdict: <n> repeats, <m> mechanical · <j> judgement. Biggest: <the one that cost most> — <fix form>.
Repo checks: <command → where it runs (hook / CI / nowhere)>
| # | Repeat (user's words) | Count | Class | Fix | Scar |
|---|-----------------------|-------|-------|-----|------|
| 1 | "<quote>" | <n>× | mechanical | <file: lint rule / hook / test> | — |
| 2 | "<quote>" | <n>× | judgement | <CLAUDE.md line / skill edit> | <quote, count> |
Proposed diffs:
<one block per row>
Incidents (once only, no fix): <quote — or none>
Count: <m> mechanical + <j> judgement = <n> repeats
Approve rows: <numbers> — nothing is written until you say so.
Do NOT pad with generic advice (every line must trace to a table row), and do NOT close with a second option for the same row.
What Claude Does / What You Do
| Claude |
You |
| Quotes and counts the repeats, reads the repo's check command, classifies each |
Say which session, and correct a misclassified row |
| Writes the check or the rule as an exact diff, with its scar |
Approve rows by number |
| Applies only the approved rows afterwards |
Decide what the team's CLAUDE.md should carry |
Related Skills
skill:audit — what skills, plugins, and hooks the workspace already has.
write:writing-skills — a skill that misfires; edit that skill rather than add a retro rule.
update-config — applying an approved hook or permission to settings.json.
1---2name: retro3effort: high4description: >-5 Use when a session's repeated mistakes or corrections should stop recurring. Triggers on: "이번 세션 회고해서 다음에 같은 실수 안 하게 해줘", "또 같은 실수네", "session retro", "stop repeating this mistake".6scenarios:7 - "Run a retro on this session so you don't make the same mistakes again"8 - "You keep making the same mistake. What should we change so it stops?"9 - "이번 세션 회고해서 다음에 같은 실수 안 하게 해줘"10 - "같은 지적을 세 번 했어. 재발 안 하게 뭘 바꿔야 해?"11 - "세션 돌아보고 CLAUDE.md나 훅에 뭘 넣을지 제안해줘"12compatibility:13 optional:14 - think-tool15 remote_mcp_note: >-16 think-tool이 있으면 반복된 실수가 검사로 잡히는 종류인지 판단의 문제인지 가르는 분류를 더 꼼꼼히 할 수 있습니다.17 Claude 설정 → MCP Servers에서 remote SSE 엔드포인트를 추가하세요.18metadata:19 source: https://skillmd.com/skills/newkayak12/retro20---2122## Standing Mandates2324- NEVER answer a mechanical repeat with a prose rule — no memory note, no CLAUDE.md line. A rule asks the same agent that just failed to remember; a check fails the build whoever forgets. Probe: a twice-corrected `@/` import and three skipped lint runs ended in two memory notes; the lint rule was only offered, `package.json` never opened.25- ALWAYS read the repo's own check command before proposing anything: `package.json` scripts, `Makefile`, CI config, git hooks. A check that exists but is unwired or skipped is the finding, not a new one.26- NEVER write or edit any file, and never save a memory, before the user approves in so many words. The run ends with the proposal. A retro that rewrites `CLAUDE.md` unasked edits a file the whole team loads.27- NEVER invent a repeat. Every item carries a quote from the session and a count; one occurrence is an `incident`, not a pattern.28- Goal: every repeat has a class, a count, and one proposed fix. The Count line recounts to the table. One pass, then wait.2930# Retro3132Reads one session's corrections, finds what repeated, and proposes the smallest change to the environment that makes it stop.3334**Not for** inventorying which skills or plugins exist (`skill:audit`) or repairing a single skill that misfires (`write:writing-skills`).3536## Process37381. **Scope.** The session the user named; none named → the current one. Source is the transcript or log the user pasted. Done when the source is identified; cannot reach it → ask once in one line.392. **Collect.** Each correction, re-ask, or "again" the user made. Quote the user's words, count occurrences, name the file or step. Fold items with the same cause into one row. Done when no two rows share a cause and each has a quote and a count.403. **Read the checks.** Open `package.json` scripts, `Makefile`, CI workflow, and `.git/hooks` or hook config; name the lint / typecheck / test command that exists and whether anything runs it automatically. No guardrail at all is itself a finding.414. **Classify** each row:42 - **Mechanical** — a fixed pattern a tool can see (a banned import shape, an unrun command, a file-location rule, an unused symbol). Default.43 - **Judgement** — cross-file consistency, naming taste, "match the surrounding style"; no check could substitute.44 A row you hesitate on is mechanical until you have tried to write the check in one sentence and failed.455. **Propose.** Mechanical → the check itself, named by file and form: a custom or built-in lint rule (`no-restricted-imports`), a pre-commit or Claude Code hook, a CI step, a test. Show the exact config diff. Judgement → one `CLAUDE.md` line or a skill edit, carrying its scar (the quote and count that earned it). Max one fix per row, no menus.466. **Stop.** Present the proposal and wait. Approval for some rows is approval for those rows only.4748## Output Template4950```51Verdict: <n> repeats, <m> mechanical · <j> judgement. Biggest: <the one that cost most> — <fix form>.5253Repo checks: <command → where it runs (hook / CI / nowhere)>5455| # | Repeat (user's words) | Count | Class | Fix | Scar |56|---|-----------------------|-------|-------|-----|------|57| 1 | "<quote>" | <n>× | mechanical | <file: lint rule / hook / test> | — |58| 2 | "<quote>" | <n>× | judgement | <CLAUDE.md line / skill edit> | <quote, count> |5960Proposed diffs:61<one block per row>6263Incidents (once only, no fix): <quote — or none>64Count: <m> mechanical + <j> judgement = <n> repeats65Approve rows: <numbers> — nothing is written until you say so.66```6768Do NOT pad with generic advice (every line must trace to a table row), and do NOT close with a second option for the same row.6970## What Claude Does / What You Do7172| Claude | You |73|--------|-----|74| Quotes and counts the repeats, reads the repo's check command, classifies each | Say which session, and correct a misclassified row |75| Writes the check or the rule as an exact diff, with its scar | Approve rows by number |76| Applies only the approved rows afterwards | Decide what the team's `CLAUDE.md` should carry |7778## Related Skills7980- `skill:audit` — what skills, plugins, and hooks the workspace already has.81- `write:writing-skills` — a skill that misfires; edit that skill rather than add a retro rule.82- `update-config` — applying an approved hook or permission to `settings.json`.