reflect

What This Skill Does

Reflect is a high-level cognitive skill for OpenClaw agents designed to facilitate autonomous self-improvement through structured conversation analysis. Its core philosophy is "Correct once, never again." By continuously monitoring interaction patterns, the agent identifies explicit user corrections, successful task resolution strategies, and implicit patterns that indicate optimal workflows. Once a learning is codified, the skill integrates these insights directly into the agent’s system instructions, personality files, or specialized skill sets. This allows the AI to evolve from a static assistant into a context-aware partner that matures in lockstep with the user's requirements.

Installation

To integrate this skill into your environment, run the following command in your terminal: clawhub install openclaw/skills/skills/stevengonsalvez/agent-reflect Once installed, you can toggle the system behavior using reflect on or reflect off. Use reflect status to check the current metrics of your agent's learning progress.

Use Cases

• Project Onboarding: After long sessions defining project requirements, use reflect to ensure all project-specific naming conventions and architectural preferences are permanently baked into the agent's context.
• Error Correction: When the agent repeatedly struggles with a specific API implementation, reflect captures the final, working solution and prevents the agent from falling into previous, buggy patterns.
• Workflow Optimization: After a successful complex deployment, trigger reflect to extract the exact steps taken, turning a one-off success into a repeatable, optimized skill.

Example Prompts

1. "Reflect on our conversation today and update my backend-developer instructions with the API error handling rules we just fixed."
2. "Reflect off; I want to experiment with some wild ideas without them being permanently encoded into my agent definition."
3. "Reflect review: Show me all the pending learnings you've identified before you commit them to the configuration files."

Tips & Limitations

• Quality Gates: The skill is designed to ignore trivial noise. It only proposes changes that are reusable, non-trivial, specific, and verified.
• Human-in-the-loop: For high-stakes changes, always use reflect review to audit proposed modifications before applying them to core architecture files.
• Confidence Thresholds: Pay close attention to HIGH confidence signals; these represent explicit user directives. LOW confidence signals should be reviewed with skepticism to avoid "poisoning" the agent's logic with one-off lucky guesses.