Reflect

Reflect

Success Criteria

1. Session summary is 3-5 bullets, not a wall of text.
2. Every identified issue cites a specific moment (not vague).
3. Every proposed improvement is actionable (starts with a verb, names a specific behavior).
4. User is explicitly asked which items to persist before anything is written.

This is a multi-turn skill. Execute each part sequentially, waiting for the user's response to each AskUserQuestion before proceeding to the next part. If a part has zero findings or proposals, skip it entirely — do not output "none found."

Part 1: Session Review

1. Summarize what was accomplished in this chat (bullet list, 3-5 items max).
2. Identify issues — any of:
   • Mistakes I made (wrong output, bad assumptions, unnecessary work)
   • Friction points (user had to repeat themselves, clarify, or correct me)
   • Unclear or verbose outputs the user had to parse
3. Propose improvements — a short numbered list of concrete changes to my behavior (e.g., "Ask before running deploy commands", "Keep plans under 50 lines"). Only include improvements that directly address identified issues; avoid generic suggestions. Each must be actionable and specific.
4. Present improvements interactively: Use AskUserQuestion with multiSelect: true. Each proposed improvement becomes one option (label = short name, description = the full actionable statement). Even if only 1 improvement is proposed, still use AskUserQuestion — include a "None of these" option so the user can decline.
5. Wait for the user's response. Persist only the selected improvements to CLAUDE.md. Do not write anything if the user selects none.

Stop here and wait for the user's AskUserQuestion response before continuing to Part 2.

Part 2: Skill Audit

Check which skills (files in .claude/skills/) were invoked or relevant during this session. If no skills were used or no changes are needed, skip this part entirely.

For each skill used:

Self-check injection

If the skill lacks Success Criteria or self-verification instructions, propose adding both (2-4 measurable criteria at the top, verification loop at the bottom).

Token efficiency

Scan the skill for redundancy, over-explanation, or sections that could be shortened without losing clarity or functionality. Focus on: duplicate examples, repeated warnings, verbose sections (>10 lines of similar advice). Propose specific edits with before/after text.

Other improvements

Note any other issues (stale references, missing edge cases, wrong defaults). Propose fixes.

Present skill changes interactively: Output the findings as text, then use AskUserQuestion with multiSelect: true. Each proposed skill change becomes one option (label = skill name + change type, description = what will be changed). Wait for the user's response, then apply only the selected changes.

Stop here and wait for the user's AskUserQuestion response before continuing to Part 3.

Part 3: Pattern Detection

If the user has asked me to do similar tasks 2+ times in this session (e.g., "format this as a table", "check types then test"), suggest creating a reusable skill for it. If no repeated patterns are found, skip this part entirely.

Present pattern proposals interactively: Describe what each proposed skill would do as text, then use AskUserQuestion with options for each proposed skill (label = skill name, description = what it automates). Include a "None" option. Wait for the user's response, then create only the confirmed skills.

Self-Verification

Before presenting output in each part, verify all Success Criteria are met. If any criterion fails, revise and re-check (up to 5 iterations). If after 5 iterations a criterion is still unmet, document which criterion failed, why, and present the best output achievable.