griffinhilly/claude-code-synthesis/retro

Retrospective

A structured review of how Claude has been performing, designed to surface patterns and turn them into feedback memories or CLAUDE.md improvements.

Scope

Parse $ARGUMENTS for scope:

Process

Step 1: Gather Evidence

Session scope: Review the current conversation — what was attempted, what succeeded, what failed, what required correction.

Weekly scope: Read MEMORY.md and PLAN.md from each active project (check the Project Dashboard in auto-memory). Look for:

• Decisions recorded in MEMORY.md
• Feedback memories in ~/.claude/projects/<project-slug>/memory/
• Recent git history (git log --since="7 days ago" --oneline) in active project directories

Project scope: Read the named project's COMP files + recent git history.

Step 2: Assess (Present to User)

Organize findings into four categories:

WORKED WELL — Things Claude did that saved time, caught errors, or produced good results. Be specific: name the task, the approach, and why it worked.

WENT POORLY — Things that wasted time, required correction, or produced bad results. Be honest. Include:

• Times the user had to correct Claude
• Approaches that failed before succeeding
• Scope creep or overengineering incidents
• Wrong assumptions that went unchecked

PATTERNS — Recurring themes across multiple instances. These are the most valuable findings. Examples:

• "Claude keeps over-engineering database queries when simple ones would work"
• "Bug fixes are faster when reproduction tests are written first"
• "Planning phases run too long before the user redirects to execution"

SUGGESTIONS — Concrete changes to try. Each should be one of:

• A new feedback memory to save
• A CLAUDE.md rule to add or modify
• A skill to create or improve
• A workflow change to try next session

Step 3: Act (With User Approval)

For each suggestion the user approves:

• Save feedback memories immediately
• Edit CLAUDE.md if warranted
• Note skill ideas in the relevant project's PLAN.md

Do NOT make changes without user approval. Present findings, wait for decisions.

Output Format

RETRO — [scope] ([date range])
 
## Worked Well
1. [specific example]
2. [specific example]
 
## Went Poorly
1. [specific example with what happened and what should have happened]
2. [specific example]
 
## Patterns
- [recurring theme]
- [recurring theme]
 
## Suggestions
1. [concrete change] → [feedback memory / CLAUDE.md edit / skill idea]
2. [concrete change] → [feedback memory / CLAUDE.md edit / skill idea]