Reflect And Learn

Reflect and Learn

I. Philosophy

Every user correction is a gift -- it reveals the gap between what the agent does and what the operator needs. But corrections are ephemeral: they live in conversation context that compacts away. This skill transforms ephemeral corrections into permanent behavioral improvements by capturing them in real-time, validating them with semantic analysis, and routing them to the right persistence target (CLAUDE.md, skill files, or AGENTS.md).

The key insight from claude-reflect: corrections made during skill execution should route back to the skill file, not just global CLAUDE.md. A correction during /deploy improves the deploy skill; a correction about code style improves the global config.

II. When to Use

• At the end of a productive session where several corrections were made.
• When the user says "remember this for next time" or "don't do that again."
• Periodically (weekly) to discover recurring workflow patterns across sessions.
• After a skill is used multiple times and the user has refined its behavior through corrections.
• When onboarding to a new project and accumulating project-specific preferences.

Do not use this skill for extracting strategic seeds from experiences (use seed-extraction). Do not use it for compressing conversation history (use compression-ritual).

III. Workflow

Step 1 -- Review captured corrections.

Examine the correction queue (populated by hooks or manual review). For each captured correction, assess:

• Confidence level (0.60-0.95): Is this a genuine correction or conversational noise?
• Scope: Global (all sessions), project-specific, or skill-specific?
• Actionability: Can this be expressed as a clear behavioral rule?

Correction signals to look for:

• Direct negations: "no, use X" / "don't use Y"
• Clarifications: "actually..." / "that's wrong"
• Explicit markers: "remember:" (highest confidence)

Step 2 -- Validate with semantic analysis.

For each candidate correction:

• Process through semantic filter to eliminate false positives
• Extract distilled, actionable statement from noisy context
• Assign final confidence score: max(regex_score, semantic_score)
• Discard anything below 0.60 confidence threshold

Step 3 -- Route to persistence target.

Determine the correct destination for each approved learning:

Step 4 -- Apply and verify.

Write the learning to the target file. Verify it does not conflict with existing rules. If a conflict exists, surface it to the user for resolution.

Step 5 -- Discover recurring patterns (optional).

Analyze session history (14+ days) for semantically similar requests:

• Group related workflows by project context
• Identify candidates for new skill creation
• Generate draft skill files with evidence trails and embedded corrections

IV. Best Practices

• Capture is automatic; application is always manual with human review.
• Route corrections to the most specific target (skill file > project CLAUDE.md > global).
• Queue must survive compaction -- back up before PreCompact events.
• Session start should display pending (unreviewed) learnings.
• After git commits, prompt for reflection (natural pause point).
• Keep confidence threshold at 0.60 minimum to filter noise.

V. Quality Checklist

• Corrections are captured with confidence scores
• Semantic validation eliminates false positives
• Routing targets the most specific persistence location
• No conflicts with existing CLAUDE.md rules
• Queue persists across session boundaries
• User reviewed and approved each applied learning

VI. Common Pitfalls

• Auto-applying without review. Regex captures sarcasm and hypotheticals.
• Only routing to global CLAUDE.md. Skill-specific corrections get diluted.
• Skipping semantic validation. False positive rate from regex alone is too high.
• Waiting too long to reflect. Correction context degrades after compaction.
• Treating affirmations as corrections. "Perfect!" is reinforcement, not change.

VII. Related Skills

• seed-extraction -- Extract strategic patterns (seeds), not behavioral corrections
• compression-ritual -- Compress conversation history into memory artifacts
• memory-garden -- Write structured memory entries for context management
• hooks-reference -- Configure capture hooks for automatic correction detection
• claude-md-guardian -- Protect CLAUDE.md integrity during learning application

Output

• One or more corrections written to the appropriate target file (skill SKILL.md, project CLAUDE.md, or global CLAUDE.md) as new behavioral rules
• A short review summary listing: corrections evaluated, confidence scores, routing decisions, and any conflicts surfaced for user resolution

Examples

Scenario 1: User says "remember — always dry-run scripts before running them" → correction assigned confidence 0.90 (explicit marker), routed to global CLAUDE.md, written as a new behavioral rule, summary returned.

Scenario 2: During a /deploy skill execution the user says "don't add the --force flag here" → correction identified as skill-specific (deploy context), routed to the deploy skill's SKILL.md, written under a "Corrections" section, summary returned.

Edge Cases

• If the correction conflicts with an existing rule in the target file, surface the conflict to the user for resolution before writing — do not overwrite silently.
• If confidence is below 0.60 (sarcasm, hypotheticals, conversational noise), discard the candidate and note it in the summary.

Anti-Patterns

• Routing all corrections to global CLAUDE.md — skill-specific corrections must go to the skill file or they get diluted across all contexts.
• Auto-applying corrections without user review — capture is automatic but application always requires explicit approval.