compound-engineering

What This Skill Does

Compound Engineering enables your AI agent to engage in self-improvement through autonomous reflection. By systematically reviewing past sessions, the skill extracts patterns, identifies potential pitfalls (gotchas), updates preferences, and archives key decisions. It fundamentally transforms your agent from a static tool into an evolving system that learns from its history. The process utilizes structured memory files, specifically MEMORY.md for long-term knowledge and dated files within the memory/ directory for granular, daily tracking. By scheduling these reviews, you ensure that your agent starts every day with the cumulative context of all previous successes and failures, effectively 'compounding' its competence over time.

Installation

To begin, ensure your OpenClaw agent environment is prepared for file system modifications. You can install the skill by executing: clawhub install openclaw/skills/skills/lxgicstudios/compound-calc. Once installed, integrate the cron functionality by adding the nightly review job to your clawdbot configuration, which ensures the agent processes the last 24 hours of logs at a designated time (e.g., 10:30 PM). If manual triggering is required, use the npx compound-engineering review command to perform an immediate audit of recent work cycles.

Use Cases

• Continuous Integration of Preferences: Automatically updating agent instructions based on subtle user feedback given throughout the week.
• Project Long-Term Memory: Maintaining a living record of complex project decisions so the agent doesn't repeat the same analytical errors on multi-day tasks.
• Error Mitigation: Logging 'gotchas' encountered during code execution to ensure the agent avoids them in future iterations.
• Workflow Optimization: Refining agent instructions in AGENTS.md whenever a faster or more accurate pattern is discovered.

Example Prompts

1. "Run a review of the last 24 hours. Extract any new patterns in how I prefer my code structured and update MEMORY.md accordingly."
2. "Perform a summary snapshot of our current progress on the database migration project and identify any unfinished tasks to add to my TODO list."
3. "Review the session from earlier today where we debugged the API connection. Log the specific cause of the failure and the solution we found into the project memory files."

Tips & Limitations

• Granularity: Keep your daily logs concise to ensure the agent can easily synthesize them during the nightly review.
• Verification: Always review the generated MEMORY.md updates once a week to ensure the agent hasn't 'over-learned' or hallucinated patterns that don't exist.
• Storage: Note that this skill requires read/write access to your local files; ensure your project directory is version-controlled with Git to easily roll back if an automated update introduces undesirable behaviors.