| name | reflect-and-learn |
| model | sonnet |
| description | Captures user corrections from a session, validates them semantically, and writes approved learnings to the most specific persistence target (skill file, project CLAUDE.md, or global CLAUDE.md). Use when: "remember this for next time", "don't do that again", "update your behavior", "capture this correction", "reflect on what we learned". |
| license | proprietary |
| category | wisdom-garden |
| triggers | ["remember this for next time","don't do that again","update your behavior","capture this correction","reflect on what we learned"] |
| tier | 1 |
| agents | ["primary"] |
| tool_dependencies | ["file_system"] |
| inputs | [{"name":"corrections","type":"string[]","description":"User corrections or learnings from the session to validate and persist","required":true}] |
| outputs | [{"name":"persisted_learnings","type":"string","description":"Approved learnings written to the most specific persistence target — skill file, project CLAUDE.md, or global CLAUDE.md"}] |
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:
