Add project-retrospective skill

[belumume]

Summary

• Adds project-retrospective skill for multi-session analysis of Claude Code projects
• Dispatches parallel opus historian agents (one per session export) with a 9-point extraction template
• Synthesizes findings into a structured analysis document with decision logs, mistake patterns, and recommendations

How it works

1. Reads session exports from ~/.claude/exports/{project-name}/
2. Spawns N background opus agents, each extracting structured data from one export
3. Combines historian reports into a 7-section analysis (vision drift, decisions, mistakes, user teaching moments, etc.)
4. Writes to docs/ and updates MEMORY.md

Prerequisites

Requires session exports created via /export. The skill gracefully handles missing exports with a clear error message.

Testing

• RED-GREEN-REFACTOR cycle with baseline comparison (agent without skill vs with skill)
• 20/20 description trigger accuracy from 4 independent evaluator agents
• Live execution test: 3 opus historians processing 68-360KB exports, producing 29 verbatim user corrections with pattern frequency analysis

Key design decisions

• Opus model for historians — exports are 50-200KB; cheaper models miss nuance
• Background Agent tool, not TeamCreate — TeamCreate has race conditions (battle-tested failure)
• "User corrections" as Reorganize the example skills #1 extraction priority — the most valuable signal for improving Claude+human collaboration
• last-N argument support — analyze only recent sessions when full history isn't needed

🤖 Generated with Claude Code