reflect（技能安全扫描）

Self-Improving Skills (Reflect)

Overview

The Reflect system enables continual learning across sessions. Instead of starting from zero every conversation, Claude learns from corrections, successful patterns, and user preferences - persisting knowledge in CLAUDE.md's Skill Memories section.

Session 1: User corrects workflow → Reflect captures learning → saves to CLAUDE.md
Session 2: Claude uses correct approach without being reminded
Session 3+: Knowledge compounds, Claude gets smarter over time

Architecture

Single Source of Truth: CLAUDE.md

All learnings are stored in your project's CLAUDE.md file under the ## Skill Memories section:

## Skill Memories

<!-- Auto-captured by SpecWeave reflect. Edit or delete as needed. -->

### Devops
- **2026-01-29**: LSP requires ENABLE_LSP_TOOL=1 env var + boostvolt/claude-code-lsps marketplace

### Frontend
- **2026-01-28**: Use shadcn/ui Button component, not custom styles
- **2026-01-27**: Prefer Vercel over Cloudflare for this Remix project

### General
- **2026-01-29**: Always check for credentials FIRST before using CLI tools

Why CLAUDE.md?

• Single file to manage (no scattered MEMORY.md files)
• Version controlled with your project
• Easy to edit, review, or delete learnings
• Automatically loaded by Claude Code

What It Learns

✅ DOES Remember

❌ Does NOT Remember

⚠️ CRITICAL: Learning Extraction Rules

This section is MANDATORY for Claude to follow when extracting learnings.

The Golden Rule

NEVER store user input verbatim. ALWAYS synthesize into actionable rules.

What Makes a Good Learning

Learning Quality Checklist (MUST PASS ALL)

Before storing ANY learning, verify:

1. ✅ Is it a complete sentence? Not truncated, not a fragment
2. ✅ Is it actionable? Contains DO/DON'T/USE/AVOID/PREFER
3. ✅ Is it specific? Names tools, patterns, files, or concepts
4. ✅ Is it understandable standalone? Someone reading it later would understand
5. ✅ Is it NOT a question? Questions are never learnings
6. ✅ Does it have context? WHY this rule exists, not just WHAT

What to REJECT (Never Store)

1. Questions - "Where should I deploy?" → NOT a learning
2. Fragments - "eplicilty how to g" → Truncated garbage
3. Raw symptoms - "always command not recognized" → No explanation
4. Duplicates - Same rule phrased differently
5. Temporary context - "for this PR", "just this time"
6. Typos/gibberish - "user pojrect", "promp"

Skill Categories

Learnings are routed to these known skills:

Usage

Manual Reflection

After completing work, manually trigger reflection:

# Reflect on current session (auto-detects skills)
/sw:reflect

# Reflect with focus prompt
/sw:reflect "Focus on the deployment patterns we discussed"

Automatic Reflection

Enable auto-reflection on session end:

# Enable automatic reflection (via stop hook)
/sw:reflect-on

# Disable automatic reflection
/sw:reflect-off

# Check reflection status
/sw:reflect-status

When enabled, the stop hook automatically:

1. Analyzes the session transcript
2. Extracts SpecWeave-specific learnings via LLM
3. Routes to appropriate skill category
4. Updates CLAUDE.md Skill Memories section

Configuration

In .specweave/config.json:

{
  "reflect": {
    "enabled": true,
    "model": "haiku",
    "maxLearningsPerSession": 3
  }
}

Disable reflection: Set "reflect": {"enabled": false}

How It Works

Extraction Flow

1. Session ends or /sw:reflect invoked
        ↓
2. Session transcript captured
        ↓
3. LLM analyzes for SpecWeave learnings
        ↓
4. Learnings validated (quality gates)
        ↓
5. Deduplicated against existing memory
        ↓
6. Written to CLAUDE.md Skill Memories

LLM Extraction

Uses Claude Haiku (fast, cheap) to extract learnings:

interface SkillLearning {
  skill: string;    // e.g., "frontend", "devops", "general"
  learning: string; // The actual learning content
}

Deduplication

Learnings are considered duplicates if:

• Content has >50% substring overlap with existing
• Same skill and semantically identical

Best Practices

For Corrections (High Signal)

Good corrections (will be captured):

"Never use that approach. Always use X because..."
"Don't create custom components. We have a design system..."
"Wrong pattern. The correct way is..."

Weak corrections (may be ignored):

"Hmm, maybe try something else?"
"That doesn't look quite right"

For Approvals (Medium Signal)

Strong approval (will be captured):

"Perfect! That's exactly how we do it."
"This is the right pattern, well done."

Neutral (won't be captured):

"OK"
"Sure"
"Proceed"

Managing Learnings

View Current Learnings

Check your CLAUDE.md file's ## Skill Memories section.

Edit Learnings

Simply edit CLAUDE.md directly. Add, modify, or delete any entry.

Clear Learnings

# Clear all learnings
/sw:reflect-clear

# Clear specific skill
/sw:reflect-clear --skill frontend

Status Check

/sw:reflect-status

Shows:

• Whether reflection is enabled
• Recent learnings
• Learning counts per skill

Troubleshooting

Learnings Not Persisting

1. Check reflection is enabled: /sw:reflect-status
2. Verify CLAUDE.md exists in project root
3. Check config: cat.specweave/config.json | grep reflect
4. Review logs: ls.specweave/logs/reflect/

Wrong Skill Detection

The LLM routes learnings based on content. If misrouted:

• Edit CLAUDE.md directly to move the learning
• Use specific language: "For frontend development..."

Garbage Learnings

If you see truncated or nonsensical entries:

1. Delete them from CLAUDE.md
2. They were likely from voice dictation or partial input

Privacy & Security

• Memory contains only patterns and learnings, not raw conversation
• No sensitive data (credentials, keys) is ever stored
• CLAUDE.md is in your repo - gitignore if needed
• Clear commands available to remove learnings

Integration with Auto Mode

When /sw:auto runs with reflection enabled:

1. Start auto session
        ↓
2. Claude executes tasks
        ↓
3. Session completes
        ↓
4. Stop hook triggers
        ↓
5. Reflect analyzes transcript
        ↓
6. CLAUDE.md updated
        ↓
7. Session ends with summary

Summary

Reflect enables correct once, apply everywhere:

1. Make correction during session
2. Reflect captures and routes to skill
3. Future sessions load from CLAUDE.md
4. Claude applies learned patterns automatically

No embeddings. No vector databases. Just your CLAUDE.md file with organized learnings.

Project-Specific Learnings

Before starting work, check for project-specific learnings:

# Check if skill memory exists for this skill
cat .specweave/skill-memories/reflect.md 2>/dev/null || echo "No project learnings yet"

Project learnings are automatically captured by the reflection system when corrections or patterns are identified during development. These learnings help you understand project-specific conventions and past decisions.