/reflect

Analyze sessions and actively improve skills based on conversation feedback.

Overview

Aspect	Details
Speed	5-15 seconds
Complexity	Medium
Purpose	Skill improvement from feedback
Version	v2.0

Usage

/reflect [skill-name]

# Examples
/reflect prompt-technical
/reflect prompt-hybrid
/reflect prompt-dotnet

How It Works

Step 1: Signal Detection

Claude scans the conversation for:

Signal Type	Confidence	Examples
Corrections	90%	"No, not like that", "I meant..."
Successes	70%	"Perfect", "Exactly", "Yes"
Edge Cases	60%	"What about...", "Can you also..."
Preferences	50%	"I prefer", "I always", "I usually"

Step 2: Analysis

Skill Reflection: prompt-technical

Signals Detected:
- Corrections: 2
- Successes: 4
- Edge Cases: 1
- Preferences: 1

Proposed Changes:

[HIGH] + Add constraint:
       "Always ask about error handling strategy"
       Evidence: User corrected twice when missing

[MED]  + Add preference:
       "Include complexity score in output"
       Evidence: User requested this consistently

[LOW]  ~ Edge case noted:
       "Multi-language prompts need detection"
       Evidence: User had Slovak/English mixed prompt

Commit message: "prompt-technical: error handling, complexity score"

Apply? [Y / n / tweaks]

Step 3: Apply Changes

If approved:

1. Modifies skill file directly
2. Records in .claude/memory/prompt-patterns.md
3. Commits changes with descriptive message

If declined:

1. Saves observations to .claude/memory/observations.md
2. Available for future review

Example Session

Scenario

You used /prompt-technical and Claude didn't ask about error handling. You corrected it twice. Later, the output was exactly what you wanted.

Running Reflect

/reflect prompt-technical

Output

Skill Reflection: prompt-technical

Signals: 2 corrections, 3 successes, 1 edge case

Proposed Changes:

[HIGH] + Constraints/NEVER:
        "Proceed without asking about error handling strategy"
        Evidence: "No, I need error handling" (2x)

[MED]  + Output Format:
        "Include estimated implementation time"
        Evidence: User asked for time estimate twice

[LOW]  ~ Note for review:
        "User prefers bullet points over paragraphs"
        Evidence: User reformatted output

Commit: "prompt-technical: require error handling, add time estimates"

Apply? [Y / n]

Applying

Y

Applied to: .claude/skills/prompt-technical/SKILL.md

Changes:
- Added constraint: Must ask about error handling
- Added output format: Include time estimate
- Saved observation: Bullet point preference

Committed: abc1234 prompt-technical: require error handling...

The skill will now:
✓ Always ask about error handling strategy
✓ Include implementation time estimates

Signal Patterns

Corrections (HIGH Priority)

Claude detects when you correct output:

"No, not like that"
"I meant..."
"Change that to..."
"Wrong, it should be..."
"Actually, I need..."

Successes (MEDIUM Priority)

Claude detects approval:

"Perfect"
"Exactly"
"Yes, that's right"
"Great"
"Looks good"

Edge Cases (MEDIUM Priority)

Claude detects missing functionality:

"What about..."
"Can you also..."
"Is there a way to..."
"What if..."

Preferences (LOW Priority, Accumulates)

Claude detects patterns over multiple sessions:

"I prefer..."
"I always..."
"I usually..."
[Repeated choices in the same direction]

Priority Levels

Priority	Meaning	Action
HIGH	Explicit correction from user	Apply immediately
MED	Repeated pattern or preference	Recommend applying
LOW	Observation for review	Save for later

Best Practices

When to Run /reflect

• After using a skill multiple times
• After a session with corrections
• When you notice patterns in your feedback
• Before ending a long session

How to Get Good Results

1. Be explicit when correcting: "No, I need X instead of Y"
2. Confirm when satisfied: "Perfect, that's exactly what I needed"
3. Mention preferences: "I prefer using TypeScript for new files"

Review Saved Observations

Periodically check .claude/memory/observations.md for accumulated patterns that haven't been applied.

Configuration

Settings in .claude/config/learning-config.json:

{
  "reflection_analysis": {
    "enabled": true,
    "min_signals_to_report": 2,
    "confidence_thresholds": {
      "correction": 0.9,
      "success": 0.7,
      "edge_case": 0.6,
      "preference": 0.5
    }
  }
}

Related Commands

• /prompt-technical - Common target for reflection
• /prompt-hybrid - Another common target
• /reflect - Run after any skill-heavy session