Reflect

Reflect - Self-Improving Memory

Analyze sessions to extract corrections and patterns, then update project skills automatically.

When to Use

• Triggered by /reflect command
• After sessions with explicit corrections or approvals
• When user provides feedback like "always do X" or "never do Y"
• Automatically via hooks at session end (if enabled)

What This Skill Does

1. Scans the conversation for learning signals:

   • Corrections: User said "don't do X" or "use Y instead"
   • Approvals: User accepted a suggestion enthusiastically
   • Patterns: Approaches that worked well repeatedly

2. Categorizes by confidence:

   • HIGH - Never do X (explicit prohibitions)
   • MEDIUM - Patterns that work well (successful approaches)
   • LOW - Observations to review later

3. Updates relevant skills:

   • ui-conventions/Skill.md - UI/UX patterns
   • code-style/Skill.md - Coding conventions
   • api-design/Skill.md - API design patterns
   • security-practices/Skill.md - Security learnings

4. Commits to Git:

   • Auto-generates descriptive commit message
   • Shows diff for review before applying
   • Allows natural language edits to proposals

How to Analyze a Session

Step 1: Identify Learning Signals

Look for patterns like:

Step 2: Map to Skills

Step 3: Propose Changes

For each learning, propose:

## Proposed Change

**Signal:** HIGH/MEDIUM/LOW
**Source:** User correction on [date]
**Learning:** [What was learned]

**Before:**
[Current content in skill]

**After:**
[Updated content with learning]

Step 4: Generate Commit Message

Format: feat(skill-name): learning description

Examples:

• feat(ui-conventions): learn to use Quasar Notify instead of alert()
• feat(code-style): learn to prefer async/await over Promise chains
• feat(api-design): learn to return 204 for DELETE endpoints

Example Session Analysis

User: Create a submit button
Claude: [Creates with inline alert()]

User: Don't use alert(). Use Quasar's $q.notify instead.
Claude: [Updates to use $q.notify]

---

/reflect triggers this skill:

Signals Detected:
1. HIGH: Never use alert() in this project
2. MEDIUM: Use Quasar's $q.notify for notifications

Skill to Update: ui-conventions

Proposed Change:
Add to "Notifications" section:
- **CRITICAL:** Never use browser alert()
- Always use: $q.notify({ type: 'positive', message: '...' })

Commit: feat(ui-conventions): learn to use Quasar Notify instead of alert()

Guidelines for Good Learning

DO:

• Be specific about what changed
• Include context (what was wrong, why it was corrected)
• Use consistent language ("never" for HIGH, "prefer" for MEDIUM)
• Link to original conversation if possible

DON'T:

• Record vague observations without context
• Store LOW confidence items that aren't useful
• Update skills without user confirmation
• Add contradictory rules

Toggle Controls

User can control this skill:

/reflect         # Trigger manual reflection
/reflect on      # Enable auto-reflection on session end
/reflect off     # Disable auto-reflection
/reflect status  # Check if auto-reflection is enabled

Files This Skill Updates

• .claude/skills/ui-conventions/Skill.md
• .claude/skills/code-style/Skill.md
• .claude/skills/api-design/Skill.md
• .claude/skills/security-practices/Skill.md

Hook Integration

When enabled via stop-hook, this skill runs automatically at session end:

1. Session ends
2. Hook triggers this skill
3. Conversation is analyzed
4. Learning signals extracted
5. Proposals generated and shown
6. User confirms/edits
7. Skills updated
8. Git commit created

References

• Video: https://youtu.be/-4nUCaMNBR8
• Skills spec: https://agentskills.io