Self-Improvement — Learn, Adapt, Compound

Three modes: Weekly Review (scheduled), Session Analysis (conversation mining), Meta-Reflection (live mistake capture). All three feed improvements back into skills, memory, and CLAUDE.md.

Mode 1: Weekly Review (Sunday 10 AM ET)

Pull this week's memory files and daily notes. Answer honestly:

1. Performance Review

• Wins: What worked? What was fast/accurate?
• Fuckups: What went wrong? Be specific. (e.g., "deployed broken f-string via SSH sed, cost 2 hours")
• Wasted effort: What took 5 attempts that should've taken 1?

2. Trade Review (if applicable)

Pull trade journal data:

• Win rate by signal type, asset, timeframe, conviction
• Biggest wins and losses (absolute + %)
• Sizing mistakes
• Signal accuracy: did high-conviction signals actually hit more?

3. Efficiency Review

Count from daily logs:

• Tool failures / retries
• Bot restarts
• Times I asked the user something I could've figured out
• Repeated tasks that should be automated

4. Skill Gaps

• What came up that I couldn't handle?
• What new tool/capability would've helped?
• What knowledge was missing?

5. Action Items

Concrete, specific improvements. Each gets owner (me or user) and deadline (this week).

6. Partner Feedback

• Review direct feedback from user this week
• What did they correct? What frustrated them?
• Incorporate into skills/workflows

Output

Save to memory as self-review-YYYY-WXX.md:

# Self-Review — Week XX, YYYY

## Wins
- ...

## Fuckups
- ...

## Wasted Effort
- ...

## Trade Performance
- Overall WR: X%
- Best signal type: X (Y% WR)
- Worst signal type: X (Y% WR)

## Efficiency
- Tool failures: X
- Bot restarts: X
- Avoidable asks: X

## Skill Gaps
- ...

## Action Items
1. [owner] action — deadline

## Partner Feedback Summary
- ...

## Top 3 Improvements This Week
1. ...
2. ...
3. ...

Mode 2: Session Analysis (Conversation Mining)

Dig through Claude Code conversation history to find systemic friction.

Phase 1: Find Conversations

# Get project directory
ls ~/.claude/projects/

# List recent conversations
ls -lt ~/.claude/projects/<encoded-path>/*.jsonl | head -20

Phase 2: Parallel Analysis

Spawn Agent subagents to analyze conversations in parallel. Each agent reads one .jsonl file and extracts:

1. What was user trying to accomplish?
2. Problems/friction that occurred (include user quotes showing frustration)
3. What worked well?
4. Repeated patterns or inefficiencies

Prioritize large files (>500KB) — they contain the meatiest sessions.

Phase 3: Synthesize

Combine findings:

• Top friction patterns ranked by frequency
• What worked well (don't lose these)
• User frustration quotes (raw evidence)

Generalize aggressively. Look for the meta-pattern behind specific issues.

Phase 4: Cross-Reference

Read current documentation:

• CLAUDE.md files (project + user level)
• All skills in ~/.claude/skills/
• Memory files

For each friction pattern, classify:

• Already fixed — note where
• Still missing — needs addition

Phase 5: Output

Create CLAUDE_IMPROVEMENTS.md with ranked suggestions. Do not apply changes — create for user review.

# Suggested Improvements

Based on analysis of N conversations from [date range].

## 1. [Issue Name]
**Problem:** [Description]
**Suggested addition to [location]:**
```markdown
[Proposed text]

Already Fixed

Potential New Skills

Raw Friction Log

• "user quote 1"
• "user quote 2"

---

## Mode 3: Meta-Reflection (Live Mistake Capture)

When a mistake happens mid-session, don't just fix it — extract the lesson:

1. **Reflect**: What went wrong and why? (Full context available now)
2. **Abstract**: What's the underlying pattern? (Not "forgot to check X" but "skipping validation before Y")
3. **Generalize**: Create a reusable rule that prevents this class of error
4. **Document**: Save as a feedback memory or update relevant skill

### Rule Writing Guidelines
- Use absolute directives: Start with "NEVER" or "ALWAYS"
- Lead with why: Explain the problem before the solution (1-3 bullets max)
- Be specific enough to act on, general enough to reuse
- Include the cost of the mistake (time, money, trust)

### Example

Bad: "Be more careful with deployments" Good: "ALWAYS validate Python syntax locally before SCP to VPS. Why: SSH sed edits broke f-strings twice (2h lost each time). How: Run python -c 'import ast; ast.parse(open(f).read())' before deploy."

---

## Integration Points
- **memory-management**: Weekly review feeds into memory curation cycle
- **know-me**: Partner feedback section feeds preference learning
- **All skills**: Action items from reviews should update specific skills