Self-Improvement Skill

IMPORTANT: This skill should be invoked AUTOMATICALLY at the end of every significant task. Do NOT ask the user — just do it silently after task completion.

When to Trigger

Invoke this skill when ANY of these happen:

User says "done", "that's it", "thanks", or moves to a new task
A QA sign-off is posted
A report is generated
An eval is completed
A Slack update is posted summarizing work
User explicitly asks to capture learnings

What to Do

Step 1: Reflect (internally, don't output to user)

Answer these questions honestly:

Where did I go wrong?
- Did I waste tokens on unnecessary queries?
- Did I ask questions I could have queried?
- Did I present wrong data as facts?
- Did I repeat the same action without progress?
- Did I miss something the user had to correct me on?
How could I have been more efficient?
- Could I have batched queries instead of sequential?
- Could I have used a skill instead of manual work?
- Did I do work the user didn't ask for?
- Did I over-engineer or under-deliver?
How to get it done faster next time?
- What was the critical path? What was wasted time?
- What tools/queries should I have used first?
- What pattern should I recognize instantly next time?
What should I never do again?
- Specific anti-patterns from this task
- Things the user explicitly corrected
What worked well?
- Approaches the user approved or praised
- Efficient patterns worth repeating

Step 2: Write to Retrospective Log

Append to ~/.claude/projects/-Users-chethanbhatbs/memory/self_retro_log.md

Format per entry:

## [DATE] — [Task Name]

**Mistakes:**
- [specific mistake with context]

**Inefficiencies:**
- [what took too long and why]

**Fixes for Next Time:**
- [concrete action item — not vague "be better"]

**What Worked:**
- [keep doing this]

**Pattern to Remember:**
- [if X happens, do Y instead of Z]

Step 3: Update Relevant Skills/Memory

If a mistake is systemic (not one-off), update the relevant skill or memory file:

Testing mistake → update /template-eval or /qa-playbook
Query mistake → update DB query references
Workflow mistake → update feedback memory files
New pattern discovered → create new feedback memory

Step 4: Check Previous Retros

Before starting any new task, scan self_retro_log.md for patterns matching the current task type. Apply lessons learned.

Anti-Patterns Log (Quick Reference)

This section gets updated over time. Check before every task.

NEVER DO

Run cron polls without a way to stop them (delete cron immediately when user says stop)
Run 3+ parallel agent-browser sessions (browser daemon deadlock)
Present partial/snapshot data as accurate totals
Query Prod DB for EPH job data (it won't be there)
Keep polling when no changes for 10+ min without suggesting alternatives
Ask user for info you can query from DB (job details, model, agent, env, ECU)
Waste tokens querying jobs the user didn't ask about

ALWAYS DO

Auto-gather context before asking questions
Suggest relevant skills inline during tasks
Run E2E tests sequentially (one at a time)
Flag data uncertainty upfront ("this is partial/estimated")
Delete cron jobs immediately when user says stop
Save results after each test (don't batch)
Check pod status before testing
Confirm environment (prod/canary/EPH) before sending HITL

self-improve