reflect-learn（反思学习）

Reflect - Self-Improvement Skill

Quick Reference

Core Philosophy

"Correct once, never again."

When users correct behavior, those corrections become permanent improvements encoded into the agent system - across all future sessions.

Workflow

Step 1: Initialize State

Check and initialize state files using the state manager:

# Check for existing state
python scripts/state_manager.py init

# State directory is configurable via REFLECT_STATE_DIR env var
# Default: ~/.reflect/ (portable) or ~/.claude/session/ (Claude Code)

State includes:

• reflect-state.yaml - Toggle state, pending reviews
• reflect-metrics.yaml - Aggregate metrics
• learnings.yaml - Log of all applied learnings

Step 2: Scan Conversation for Signals

Use the signal detector to identify learnings:

python scripts/signal_detector.py --input conversation.txt

Signal Confidence Levels

See signal_patterns.md for full detection rules.

Step 3: Classify & Match to Target Files

Map each signal to the appropriate target:

Learning Categories:

See agent_mappings.md for mapping rules.

Step 4: Check for Skill-Worthy Signals

Some learnings should become new skills rather than agent updates:

Skill-Worthy Criteria:

• Non-obvious debugging (>10 min investigation)
• Misleading error (root cause different from message)
• Workaround discovered through experimentation
• Configuration insight (differs from documented)
• Reusable pattern (helps in similar situations)

Quality Gates (must pass all):

• [ ] Reusable: Will help with future tasks
• [ ] Non-trivial: Requires discovery, not just docs
• [ ] Specific: Can describe exact trigger conditions
• [ ] Verified: Solution actually worked
• [ ] No duplication: Doesn't exist already

See skill_template.md for skill creation guidelines.

Step 5: Generate Proposals

Produce output in this format:

# Reflection Analysis

## Session Context
- **Date**: [timestamp]
- **Messages Analyzed**: [count]
- **Focus**: [all agents OR specific agent name]

## Signals Detected

| # | Signal | Confidence | Source Quote | Category |
|---|--------|------------|--------------|----------|
| 1 | [learning] | HIGH | "[exact words]" | Code Style |
| 2 | [learning] | MEDIUM | "[context]" | Architecture |

## Proposed Agent Updates

### Change 1: Update [agent-name]

**Target**: `[file path]`
**Section**: [section name]
**Confidence**: [HIGH/MEDIUM/LOW]
**Rationale**: [why this change]

--- a/path/to/agent.md +++ b/path/to/agent.md @@ -82,6 +82,7 @@ ## Section

* Existing rule +* New rule from learning

## Proposed New Skills

### Skill 1: [skill-name]

**Quality Gate Check**:
- [x] Reusable: [why]
- [x] Non-trivial: [why]
- [x] Specific: [trigger conditions]
- [x] Verified: [how verified]
- [x] No duplication: [checked against]

**Will create**: `.claude/skills/[skill-name]/SKILL.md`

## Conflict Check

- [x] No conflicts with existing rules detected
- OR: Warning - potential conflict with [file:line]

## Commit Message

reflect: add learnings from session [date]

Agent updates:

• [learning 1 summary]

New skills:

• [skill-name]: [brief description]

Extracted: [N] signals ([H] high, [M] medium, [L] low confidence)

## Review Prompt

Apply these changes?
- `Y` - Apply all changes and commit
- `N` - Discard all changes
- `modify` - Adjust specific changes
- `1,3` - Apply only changes 1 and 3
- `s1` - Apply only skill 1
- `all-skills` - Apply all skills, skip agent updates

Step 6: Handle User Response

On Y (approve):

1. Apply each change using Edit tool
2. Run git add on modified files
3. Commit with generated message
4. Update learnings log
5. Update metrics

On N (reject):

1. Discard proposed changes
2. Log rejection for analysis
3. Ask if user wants to modify any signals

On modify:

1. Present each change individually
2. Allow editing the proposed addition
3. Reconfirm before applying

On selective (e.g., 1,3):

1. Apply only specified changes
2. Log partial acceptance
3. Commit only applied changes

Step 7: Update Metrics

python scripts/metrics_updater.py --accepted 3 --rejected 1 --confidence high:2,medium:1

Toggle Commands

Enable Auto-Reflection

/reflect on
# Sets auto_reflect: true in state file
# Will trigger on PreCompact hook

Disable Auto-Reflection

/reflect off
# Sets auto_reflect: false in state file

Check Status

/reflect status
# Shows current state and metrics

Review Pending

/reflect review
# Shows low-confidence learnings awaiting validation

Output Locations

Project-level (versioned with repo):

• .claude/reflections/YYYY-MM-DD_HH-MM-SS.md - Full reflection
• .claude/reflections/index.md - Project summary
• .claude/skills/{name}/SKILL.md - New skills

Global (user-level):

• ~/.claude/reflections/by-project/{project}/ - Cross-project
• ~/.claude/reflections/by-agent/{agent}/learnings.md - Per-agent
• ~/.claude/reflections/index.md - Global summary

Memory Integration

Some learnings belong in auto-memory (~/.claude/projects/*/memory/MEMORY.md) rather than agent files:

When a signal is LOW confidence and project-specific, prefer writing to MEMORY.md over modifying agents.

Safety Guardrails

Human-in-the-Loop

• NEVER apply changes without explicit user approval
• Always show full diff before applying
• Allow selective application

Git Versioning

• All changes committed with descriptive messages
• Easy rollback via git revert
• Learning history preserved

Incremental Updates

• ONLY add to existing sections
• NEVER delete or rewrite existing rules
• Preserve original structure

Conflict Detection

• Check if proposed rule contradicts existing
• Warn user if conflict detected
• Suggest resolution strategy

Integration

With /handover

If auto-reflection is enabled, PreCompact hook triggers reflection before handover.

With Session Health

At 70%+ context (Yellow status), reminders to run /reflect are injected.

Hook Integration (Claude Code)

The skill includes hook scripts for automatic integration:

# Install hook to your Claude hooks directory
cp hooks/precompact_reflect.py ~/.claude/hooks/

Configure in ~/.claude/settings.json:

{
  "hooks": {
    "PreCompact": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "uv run ~/.claude/hooks/precompact_reflect.py --auto"
          }
        ]
      }
    ]
  }
}

See hooks/README.md for full configuration options.

Portability

This skill works with any LLM tool that supports:

• File read/write operations
• Text pattern matching
• Git operations (optional, for commits)

Configurable State Location

# Set custom state directory
export REFLECT_STATE_DIR=/path/to/state

# Or use default
# ~/.reflect/ (portable default)
# ~/.claude/session/ (Claude Code default)

No Task Tool Dependency

Unlike the previous agent-based approach, this skill executes directly without spawning subagents. The LLM reads SKILL.md and follows the workflow.

Git Operations Optional

Commits are wrapped with availability checks - if not in a git repo, changes are still saved but not committed.

Troubleshooting

No signals detected:

• Session may not have had corrections
• Try /reflect review to check pending items

Conflict warning:

• Review the existing rule cited
• Decide if new rule should override
• Can modify before applying

Agent file not found:

• Check agent name spelling
• Use /reflect status to see available targets
• May need to create agent file first

File Structure

reflect/
├── SKILL.md                      # This file
├── scripts/
│   ├── state_manager.py          # State file CRUD
│   ├── signal_detector.py        # Pattern matching
│   ├── metrics_updater.py        # Metrics aggregation
│   └── output_generator.py       # Reflection file & index generation
├── hooks/
│   ├── precompact_reflect.py     # PreCompact hook integration
│   ├── settings-snippet.json     # Settings.json examples
│   └── README.md                 # Hook configuration guide
├── references/
│   ├── signal_patterns.md        # Detection rules
│   ├── agent_mappings.md         # Target mappings
│   └── skill_template.md         # Skill generation
└── assets/
    ├── reflection_template.md    # Output template
    └── learnings_schema.yaml     # Schema definition