Reflect Skill
Analyze conversations to extract corrections, preferences, and best practices, then persist them to skill files. Supports manual triggering, automatic hooks, and Git versioning.
Commands
| Command | Description |
|---|---|
/reflect | Analyze current session and propose skill updates |
/reflect [skill-name] | Reflect and update a specific skill |
reflect on | Enable automatic reflection via stop hook |
reflect off | Disable automatic reflection |
reflect status | Show current reflection mode |
Manual Reflection Flow
1. Analyze Session for Signals
Scan the conversation and categorize by confidence:
HIGH Confidence - Explicit corrections
- "No, use X instead of Y"
- "Never [action] in this project"
- "Always check for [condition]"
MEDIUM Confidence - Success patterns
- Approaches that worked after iteration
- Patterns user approved or praised
LOW Confidence - Observations
- Implied preferences
- Patterns to review later
2. Present Review & Approval
📊 Signals Detected:
HIGH CONFIDENCE:
- "Never generate button styles - reference existing components"
- "Always check for SQL injections"
MEDIUM CONFIDENCE:
- Arrow functions preferred for callbacks
LOW CONFIDENCE:
- Prefers concise explanations
📝 Proposed Changes to [skill-name].md:
+ ## Security
+ - Always validate database queries for SQL injection
+
+ ## Code Style
+ - Never generate UI components; reference existing library
💬 Commit message: "Add SQL injection and UI component rules"
Accept changes? [Y/n/edit with natural language]User options:
Y- Accept all changesn- Reject- Natural language: "Remove the arrow function one" or "Change the wording to..."
3. Update Skill & Commit
- Locate target skill file (project or
~/.claude/skills/) - Merge new learnings (newer rules override conflicts)
- If Git repo: stage, commit with message, push
Automatic Reflection
Toggle Commands
reflect on # Enable auto-reflection on session end
reflect off # Disable auto-reflection
reflect status # Check current modeHow It Works
When enabled, the stop hook triggers automatic learning:
- Session ends → stop hook fires
- Analyzes conversation for learnings
- Updates skill file (shows notification: "Learned from session: [skill-name]")
- Commits to Git if configured
Hook Configuration
The session-end hook is auto-registered when you install this plugin. No manual configuration needed.
State stored in ~/.claude/.reflect-state (enabled/disabled).
Git Integration
Version your skills to track evolution over time:
cd ~/.claude/skills
git init
git remote add origin <your-repo>Benefits:
- See how skills evolve with each learning
- Rollback if a learning causes regressions
- Track system improvement over time
Learning Categories
| Category | Examples |
|---|---|
| Code Style | Naming conventions, formatting |
| Security | Input validation, SQL injection |
| Testing | Coverage requirements, edge cases |
| API Design | Endpoint conventions, responses |
| UI/UX | Component usage, accessibility |
| Workflow | PR process, commit format |
Example
You: Review the auth module
Claude: [Reviews, misses SQL injection]
You: You missed SQL injection - always check for that
You: /reflect
Claude: 📊 HIGH: "Always check for SQL injections"
📝 Proposed for security-review.md:
+ - Validate all database queries for SQL injection
Accept? [Y/n]
You: Y
Claude: ✅ Updated security-review.md
📤 Committed: "Add SQL injection validation rule"Next session: Claude automatically checks for SQL injection.
Goal
Correct once, never again. Build a self-improving system where learnings persist and compound over time.