Sw Self Improving Agent

What This Skill Does

The Sw Self Improving Agent skill provides a robust architectural framework for building agents that evolve through continuous feedback. By leveraging a centralized RULES.md file, this skill enables your OpenClaw agent to institutionalize user corrections, transforming ephemeral user feedback into persistent, atomic operating guidelines. Instead of relying solely on the agent's initial prompt, this skill forces the agent to treat every user correction as a permanent update to its behavioral pattern, effectively creating a feedback loop that minimizes repetitive errors and optimizes communication styles over time.

Installation

To integrate this skill into your environment, execute the following command in your terminal:

clawhub install openclaw/skills/skills/amdf01-debug/sw-self-improving-agent

Once installed, ensure your project directory contains a RULES.md file and update your AGENTS.md file with the integration snippet provided in the skill documentation to trigger the automated logging process.

Use Cases

This skill is ideal for power users who find themselves repeatedly correcting their agents on trivial preferences or specific procedural workflows. It is particularly effective for teams managing complex, long-running projects where specific business logic must be preserved across multiple sessions. Use it to enforce communication preferences, such as brevity or tone, and to mandate operational checks, like verifying file status before execution.

Example Prompts

1. "You keep including pleasantries in your responses, which I don't want. Please log this as a rule: always provide just the raw output without any introductory or concluding text."
2. "My agent keeps suggesting meeting times without checking my calendar first. Let's add a rule to the self-improvement loop: always verify availability in the calendar file before proposing a slot."
3. "Whenever you reference a project, you're missing the current status. Add a rule to the system: always pull the latest status update from the /projects directory when mentioning a project by name."

Tips & Limitations

To ensure optimal performance, keep your RULES.md file under 150 entries to avoid model degradation. Atomic, single-sentence rules are the most effective. Periodically audit your ruleset to prune obsolete entries or resolve contradictions. Remember, this skill is a pattern-matching tool; focus on writing generalized rules (e.g., 'Always verify X') rather than instance-specific notes to maximize the model's predictive accuracy over time.