reflection（reflection 分析）

Agents repeat mistakes. Not because they're incapable — because they forget. This skill changes that. Your agent pauses before delivering, catches its own blind spots, and remembers lessons for next time.

When to Use

User needs quality assurance beyond "looks good to me." Agent handles pre-delivery evaluation, post-mistake analysis, pattern detection across sessions, and proactive lesson surfacing before repeating errors.

How It Works

         ┌──────────────────────────────────────────────┐
         │              SELF REFLECTION LOOP            │
         └──────────────────────────────────────────────┘
                              │
         ┌────────────────────┼────────────────────┐
         ▼                    ▼                    ▼
    ┌─────────┐         ┌──────────┐         ┌─────────┐
    │  PRE    │         │  POST    │         │PATTERN  │
    │DELIVERY │         │ MISTAKE  │         │DETECTED │
    └────┬────┘         └────┬─────┘         └────┬────┘
         │                   │                    │
         │  "Before I send   │  "User corrected   │  Same mistake
         │   this, let me    │   me. Why?"        │  3 times...
         │   double-check"   │                    │
         │                   │                    │
         └───────────────────┴────────────────────┘
                              │
                              ▼
                    ┌─────────────────┐
                    │ 7-DIMENSION     │
                    │ EVALUATION      │
                    │ (30 seconds)    │
                    └────────┬────────┘
                              │
              ┌───────────────┴───────────────┐
              ▼                               ▼
       ┌─────────────┐                 ┌─────────────┐
       │  ALL CLEAR  │                 │ ISSUE FOUND │
       │  Deliver    │                 │ Fix first   │
       └─────────────┘                 └──────┬──────┘
                                              │
                                              ▼
                                    ┌─────────────────┐
                                    │  LOG LESSON     │
                                    │  Miss → Root    │
                                    │  → Prevention   │
                                    └────────┬────────┘
                                              │
                                              ▼
                                    ┌─────────────────┐
                                    │  INJECT NEXT    │
                                    │  TIME           │
                                    │  "Before we     │
                                    │   proceed..."   │
                                    └─────────────────┘

The Three Triggers

1. 🔍 Pre-Delivery

Before sending important work, pause. 30 seconds. Quick scan of 7 dimensions.

When: Code, architecture, strategy, any deliverable the user will act on.

2. ❌ Post-Mistake

User corrected you. That's data. Capture it before the session ends.

When: User says "actually...", "no, that's wrong", "I meant...", frustration signals.

3. 🔄 Pattern Detection

Same category appearing 3+ times? That's not coincidence — it's a blind spot.

When: After logging 5 reflections, weekly review, or heartbeat trigger.

Architecture

Memory lives in ~/reflection/. See memory-template.md for setup.

~/reflection/
├── memory.md           # Status + preferences + stats
├── reflections.md      # Log (most recent first)
├── patterns.md         # Detected patterns
└── archive/            # Monthly archives

Quick Reference

Core Rules

1. Check Lessons Before Acting

Before significant work, scan ~/reflection/patterns.md. Surface relevant lessons:

"Before we proceed — I have a lesson from past work on [topic]: [summary]."

2. Use 7-Dimension Evaluation

If any dimension scores below 7/10 → fix before delivering.

3. Log Every Correction

When user corrects you:

1. STOP and acknowledge
2. Analyze root cause
3. Log to ~/reflection/reflections.md:

## YYYY-MM-DD | [category]
**Miss:** What went wrong
**Root:** Why (5 whys)
**Fix:** Prevention rule

4. Detect Patterns (After 5 Reflections)

• Same category 3+ times → create prevention rule
• Same mistake twice → escalate to pattern
• Improvement trend → document what worked

5. Categories for Every Reflection

Default: technical, communication, assumptions, process, scope

6. Archive Monthly

Move processed reflections to ~/reflection/archive/YYYY-MM.md. Keep reflections.md lean.

7. Track Streaks

Days since repeated mistake. Resets on pattern recurrence. Celebrate milestones.

Pattern Lifecycle

┌──────────────┐     ┌──────────────┐     ┌──────────────┐     ┌──────────────┐
│   EMERGING   │ ──▶ │    ACTIVE    │ ──▶ │  MONITORING  │ ──▶ │   RESOLVED   │
│  2 similar   │     │  3+ times    │     │  Prevention  │     │  30 days     │
│  reflections │     │  → create    │     │  in place    │     │  clean       │
└──────────────┘     │    rule      │     └──────────────┘     └──────────────┘
                     └──────────────┘

Patterns in ~/reflection/patterns.md:

## [Pattern Name]
category: technical
frequency: 4 occurrences
status: active | monitoring | resolved

**Pattern:** What keeps happening
**Root:** Why this pattern exists
**Prevention:** Rule to break it
**Last seen:** YYYY-MM-DD
**Streak:** X days without recurrence

The "Inject Next Time" Superpower

The skill's real value: surfacing lessons BEFORE you repeat mistakes.

How it works:

1. Before starting work, identify task domain
2. Check ~/reflection/patterns.md for active patterns
3. If relevant pattern exists → mention it naturally

Example:

"Before we build this API — I have a lesson about timeout handling from a previous project. Let me make sure to include proper error timeouts this time."

Setup

On first use, read setup.md for integration guidelines. Creates memory files in ~/reflection/ (user is informed where data is stored if they ask).

Common Traps

Related Skills

Install with clawhub install <slug> if user confirms:

• memory — persistent memory patterns
• decide — decision-making autonomy
• learning — adaptive learning system

Feedback

• If useful: clawhub star reflection
• Stay updated: clawhub sync