reflect

Version: 1.0.0 | Category: tooling | Loop: 1.5

Correct once, never again. Transform ephemeral session corrections into persistent skill improvements.

Quick Start

# Manual reflection after a session
/reflect

# Target a specific skill
/reflect code-review

# Enable automatic reflection on session end
/reflect-on

# Disable automatic reflection
/reflect-off

# Check current status
/reflect-status

What It Does

The Reflect skill solves a fundamental LLM limitation: sessions don't persist learning. When you correct Claude's output, that correction is forgotten by the next session.

Reflect captures:

• Corrections: "No, use X instead" -> HIGH confidence learning
• Explicit Rules: "Always do X" -> HIGH confidence learning
• Approvals: "Perfect, that's right" -> MEDIUM confidence learning
• Observations: Implicit preferences -> LOW confidence learning

And writes them to skill files where they become permanent knowledge.

How It Works

Session -> Corrections/Patterns -> /reflect -> Skill Updated -> Next Session Improved

The 7-Phase Pipeline

1. Signal Detection - Scan conversation for learning signals
2. Skill Mapping - Map signals to invoked skills
3. Confidence Classification - Apply VERIX-aligned confidence levels
4. Change Proposal - Generate diff preview
5. Apply Updates - Update SKILL.md files
6. Memory Storage - Store in Memory MCP for Meta-Loop
7. Git Commit - Version the evolution

Confidence Levels

| Level | Confidence | Triggers | |-------|------------|----------| | HIGH | 0.90 | "Always", "Never", explicit corrections | | MEDIUM | 0.75 | Approvals, patterns (2+ occurrences) | | LOW | 0.55 | Single observations, style cues |

Commands

| Command | Description | |---------|-------------| | /reflect | Manually trigger reflection on current session | | /reflect [skill] | Reflect on specific skill only | | /reflect-on | Enable automatic reflection on session end | | /reflect-off | Disable automatic reflection | | /reflect-status | Show current toggle state |

Example Output

## Session Reflection Report

### Signals Detected
- 2 corrections (HIGH)
- 1 approval (MEDIUM)

### Proposed Updates

**Skill: debug** (v2.1.0 -> v2.1.1)

```diff
+ ### High Confidence [conf:0.90]
+ - ALWAYS check for null before accessing properties [ground:user-correction:2026-01-05]

[Y] Accept [N] Reject [E] Edit

## Skill File Format

Reflect adds a `LEARNED PATTERNS` section to skill files:

```markdown
## LEARNED PATTERNS

### High Confidence [conf:0.90]
- ALWAYS check for SQL injection [ground:user-correction:2026-01-05]

### Medium Confidence [conf:0.75]
- Prefer async/await over .then() [ground:approval-pattern:2-sessions]

### Low Confidence [conf:0.55]
- User may prefer verbose logging [ground:observation:1-session]

Integration

• Memory MCP: Stores learnings for Meta-Loop aggregation
• Skill Forge: Uses safe update patterns for SKILL.md
• Git: Versions skill evolution with descriptive commits
• Meta-Loop: Aggregated learnings feed 3-day optimization

Safety Rules

1. NEVER modifies eval-harness (frozen per Bootstrap Loop)
2. ALWAYS shows diff preview before applying
3. HIGH confidence changes require explicit approval
4. Auto-apply only for MEDIUM/LOW when reflect-on is active

Architecture: Loop 1.5

Reflect implements Loop 1.5 - the missing feedback layer:

Loop 1:   Execution (per-request)
               |
Loop 1.5: Reflect (per-session) <-- THIS SKILL
               |
Loop 2:   Quality (per-session)
               |
Loop 3:   Meta-Loop (3-day)

Files

skills/tooling/reflect/
  SKILL.md        # Main skill definition
  README.md       # This file
  examples/       # Usage examples
  tests/          # Validation tests
  resources/      # Templates

See Also

• /skill-forge - Safe skill file updates
• /prompt-architect - Constraint classification
• /bootstrap-loop - Meta-skill improvement
• /memory-manager - Memory MCP operations

---

Confidence: 0.85 (ceiling: observation 0.95)