Your First Custom Agent

Here's where OpenClaw becomes yours. The file agents.md is your agent's system prompt — the instructions that define who it is, what it does, how it communicates, and what boundaries it operates within. Every interaction your agent has starts by reading this file. It's the most important file in your workspace.

A vague agents.md produces a vague agent. "You are a helpful assistant" gives you generic ChatGPT behavior. That's not what we're building. We're building an agent with a specific role, clear boundaries, a defined communication style, and a structured output format. The difference in quality is night and day.

Start with the role. Be specific. Not "you help with coding" but "you are a TypeScript developer specializing in Next.js applications. You write tests before implementation. You use bun, never npm." The more specific the role, the more focused and useful the agent becomes. Generic agents produce generic work.

Boundaries matter more than people realize. An agent without boundaries will try to do everything — and do nothing well. Tell it what it should NOT do. "Do not modify files outside of the /src directory." "Do not run destructive git commands without explicit approval." "Do not install new dependencies without asking first." These constraints aren't limitations; they're what make your agent trustworthy.

Output format is the secret weapon most people miss. Define exactly how your agent should respond. Should it use markdown headers? Bullet points? Code blocks with language annotations? A specific template? When you tell an agent "always respond with a Summary section, then Changes Made, then Next Steps," every response becomes predictable, scannable, and useful. Consistency is the hallmark of a well-configured agent.