team-skill-maintainer

name	team-skill-maintainer
description	Use when retrospecting on AI collaboration, recording bad cases, updating skill rules, or managing skill versions.

Team Skill Maintainer

Core Principle

Every AI collaboration failure is an opportunity to improve skills. Bad Cases MUST be recorded in the affected skill's badCases/ directory and rules fed back into SKILL.md.

When to Use

Use this skill when:

• Retrospective analysis in /team-retro
• Recording bad cases
• Suggesting skill rule updates
• Managing skill versions

Do NOT use this skill when:

• Creating new skills from scratch (use superpowers:writing-skills)
• Implementing code or fixing bugs (use team-implementation-guard)
• Running a retro on a change that hasn't been archived — archive first with /team-archive

Retrospective Process

1. Review change artifacts: proposal, tasks, verification records
2. Analyze collaboration: which phases went well? which had issues?
3. Identify failures: AI skipped steps? rules unclear?
4. Record bad case: write to affected skill's badCases/ directory (NOT inline in SKILL.md)
5. Update rules: add new rules to SKILL.md — only rules, not case details

Non-Negotiable Rules

• Do NOT hide failures — even embarrassing ones
• Do NOT complain without fixing — every bad case MUST have a suggested rule adjustment
• Do NOT repeat the same mistake — update skills after retro
• Do NOT embed bad cases in SKILL.md — write to badCases/ directory, keep SKILL.md lean

Skill Update Principles

1. Specific: new rules must describe specific scenarios and behaviors
2. Actionable: AI can understand and execute
3. Verifiable: can be validated by test scenarios
4. Minimal: only add necessary rules, don't over-restrict

Version Management

After each skill update:

1. Record update date and changes
2. Reference the update in the retro report
3. Verify new rules prevent the previous failure

Bad Case Format

Write bad cases to the affected skill's badCases/ directory:

skills/<skill-name>/badCases/<NNN>-<short-name>.md

File naming: sequential number + short descriptive name (e.g., 001-stop-between-tasks.md).

# Bad Case NNN: <one-line summary>

- **Date**: ...
- **Change**: <change-id>
- **Phase**: propose / plan / apply / verify / review / archive

## Trigger
User said:
> [exact quote]

## Wrong Behavior
[What the AI did that it shouldn't have]

## Expected Behavior
[What the AI should have done]

## Root Cause
[Why the skill failed to prevent this]

## Rule Changes
**Affected Skill**: [skill-name]

New/modified rules:
1. [Specific rule 1]
2. [Specific rule 2]

Example:

skills/team-implementation-guard/badCases/001-stop-between-tasks.md

# Bad Case 001: AI Paused Between Tasks Asking for Confirmation

- **Date**: 2026-05-31
- **Change**: chunked-upload
- **Phase**: apply

## Trigger
User invoked `/team-apply`. AI completed Task 1.1 and ended with "开始 Task 1.2?"

## Wrong Behavior
After each task, AI reported progress then asked "开始 Task X.Y?" instead of continuing.

## Expected Behavior
AI should immediately proceed to the next task without asking.

## Root Cause
Task Loop rule said "ONE TASK AT A TIME" but never stated the loop should continue autonomously.

## Rule Changes
**Affected Skill**: team-implementation-guard

1. Task Loop preamble: "autonomous — immediately proceed to the next"
2. Task Loop step 8: "Immediately continue — do NOT ask for permission"
3. Common Mistakes table: added "Pause between tasks" entry

Common Failure Patterns

Pattern	Suggested Rule
User pushes to skip process	Add "refuse to skip" rule
AI persuaded to change behavior	Add "non-negotiable" checklist
Rule description too vague	Add specific examples
New scenario not covered	Add checklist item
Missing verification steps	Add layered verification rules

Retro Checklist

• Reviewed collaboration records from all phases
• Identified all failures (≥ 0)
• Each failure has scenario, wrong behavior, expected behavior
• Each failure has suggested rule adjustment
• Bad case written to affected skill's badCases/ directory
• Only new rules added to SKILL.md (no case details)
• Output retro report
• Verified whether test scenarios need updating

Bad Cases

When this skill fails, record in badCases/ directory:

• Input: what retro or skill update was requested
• Wrong output: what the AI did that it shouldn't have (e.g., hid a failure, skipped a rule update)
• Expected output: what it should have done
• New rule needed: how to prevent recurrence

See badCases/ for case history.