Rewrite retrospective skill to require structural countermeasures

[ikuwow]

Summary

Rewrite the /retrospective skill to produce actionable, structural countermeasures instead of vague improvement suggestions.

The previous version tended to output compliance-based conclusions like "follow the existing rule more carefully" or "pay more attention," which provide no real prevention. The new version requires every countermeasure to be an external enforcement mechanism (rule addition, hook, linter, script, etc.) and explicitly prohibits willpower-based fixes.

What changed

• Replaced the 4 abstract analysis categories (communication issues, improvements, rule proposals, project improvements) with a 2-step format: list problem behaviors, then propose structural countermeasures
• Added "Prohibited Countermeasures" section banning compliance/attention/memory-based fixes
• Added "Prohibited Root Causes" section banning "rule was not followed" as a root cause
• Enforced compact output format (problem-countermeasure pairs only)

Motivation

In practice, the retrospective output was too long and ended up attributing issues to the model not following existing rules. This is not useful because:

1. The model cannot reliably improve its own attention or compliance through self-reflection
2. If a rule exists but the model violated it, the rule or its enforcement is the problem, not the model's diligence

Verification

• Confirmed the file passes pre-commit hooks