Reflect

Reflect

Extract learnings from sessions and persist to skill files.

When to Reflect

Run /reflect after sessions with:

• Corrections - "actually use X", "no, do it this way"
• Discoveries - patterns that worked well
• Failures - approaches that didn't work

Usage

/reflect              # Analyze current session, suggest skill updates
/reflect skill-name   # Target specific skill for updates

Note: This is manual-only. Run before ending sessions with learnings.

Why Manual-Only?

Auto-detection via Stop hooks doesn't work reliably:

Stop hooks fire but can't communicate back - making them useless for reminders. Other self-improving skill implementations (autoskill, reflect-skill) also use manual triggering for this reason.

Process

1. Source - Determine conversation source (see Data Sources below) 2. Analyze - Find corrections, successes, patterns 3. Classify - High/Medium/Low confidence learnings 4. Propose - Show suggested skill updates 5. Approve - Wait for user confirmation before writing

Data Sources

Try sources in order, use first available:

1. ccrecall.db (Full History)

If user has ccrecall + mcp-sqlite-tools:

SELECT timestamp, role, content FROM messages
WHERE session_id = (SELECT MAX(session_id) FROM sessions)
ORDER BY timestamp DESC LIMIT 100;

2. In-Context (Current Session)

Fallback when ccrecall unavailable:

• Analyze conversation visible in current context window
• Limited to ~100k tokens of recent history
• Still effective for single-session learnings

Note which mode is active:

[reflect] Using: ccrecall.db (full history)
-- or --
[reflect] Using: in-context (current session only)

Destination Logic

• In repo with .claude-plugin/ → update skill in-place
• Otherwise → prompt: project .claude/skills/ or global ~/.claude/skills/

References

• analysis-patterns.md - Pattern detection rules

Analysis Patterns

Rules for detecting learnings in conversation history.

Correction Patterns (High Confidence)

Direct user corrections indicate wrong assumptions:

Success Patterns (Medium Confidence)

Indicators of good approaches:

Context Patterns (Low Confidence)

Project-specific context worth noting:

Data Source Detection

Check for ccrecall availability first:

1. Check if mcp-sqlite-tools is available
2. Try to open ~/.claude/ccrecall.db
3. If success → use ccrecall mode
4. If fail → use in-context mode

ccrecall Query Template

SELECT timestamp, role, content
FROM messages
WHERE session_id = (SELECT MAX(session_id) FROM sessions)
ORDER BY timestamp DESC
LIMIT 100;

In-Context Analysis

When ccrecall unavailable, scan the current conversation context for:

• User messages containing correction patterns
• Assistant messages that were subsequently corrected
• Sequences showing: proposal → correction → revision

Classification Rules

High Confidence

• Explicit correction with clear alternative
• User uses strong language ("always", "never", "must")
• Same correction repeated in session

Medium Confidence

• Implicit preference (user chooses one option)
• Success without explicit praise
• Pattern works but not explicitly validated

Low Confidence

• Context mentioned but not corrected
• Single occurrence
• Ambiguous preference

Output Format

When proposing updates:

## Proposed Learning

**Confidence**: High/Medium/Low
**Source**: [quote from conversation]
**Learning**: [extracted rule]
**Target Skill**: [skill name or "new skill"]

### Suggested Addition

[markdown to add to skill file]