agent-builder

What This Skill Does

The agent-builder skill is the foundational architect for your OpenClaw ecosystem. It transforms abstract ideas into high-performing, structured, and compliant agents. By orchestrating the creation of core workspace files—including SOUL.md, IDENTITY.md, AGENTS.md, USER.md, and HEARTBEAT.md—this skill ensures that your agent is built on a solid, repeatable architectural framework. It follows a rigorous four-phase workflow: interview, generation, guardrail enforcement, and acceptance testing, ensuring that your agent is not only functional but also aligned with your specific risk appetite and behavioral requirements.

Installation

To install this skill, run the following command in your terminal: clawhub install openclaw/skills/skills/icemastert/agent-builder-1-0-0

Use Cases

1. Greenfield Agent Development: Use this when starting a project from scratch to ensure all necessary system files are generated with best-practice defaults.
2. Behavioral Calibration: If an existing agent is too chatty, drifting in its persona, or failing to adhere to your safety guardrails, use agent-builder to iterate on its SOUL.md and AGENTS.md files.
3. Safety Hardening: Easily update your agent's autonomy levels and destructive action protocols to prevent overreach in production environments.
4. Scalability: Standardize the structure of multiple agents across different workspaces, ensuring consistent performance and maintenance patterns.

Example Prompts

1. "I want to build a personal research assistant for Telegram. It should have operator-level autonomy and never post messages to public groups without my explicit confirmation. Start by asking me the necessary questions."
2. "My current agent is being too helpful and keeps trying to execute shell commands when I ask for advice. Let's iterate on its AGENTS.md to add a stricter 'think-before-action' guardrail."
3. "Review my current agent's SOUL.md and HEARTBEAT.md. I want it to be more concise and I need a better heartbeat plan for checking long-running tasks once every hour."

Tips & Limitations

• Be Specific: The quality of the generated workspace depends heavily on your answers during the interview phase. Provide concrete examples of 'fail states' to help the agent define its guardrails better.
• Manual Review: While agent-builder generates robust templates, always conduct a final manual review of AGENTS.md to confirm that critical business logic and safety prohibitions are correctly represented.
• Iteration Mindset: Treat your agent as a living system. Use the iteration workflow regularly to refine its performance based on real-world usage data and failure modes observed over time.