How to brief an AI coding agent
Tools like Claude Code do not just answer, they act: they read your files, run commands, and edit code. That changes how you should write the brief.
A chat model gives you text; an agentic coding tool takes actions in your project. The prompt is less “question” and more “work order”, and the difference shows up in four places.
State the outcome, not the steps
Bad: “open the config file and change line 40”. Good: “requests to /api/export time out after 30 seconds; make exports reliable for files up to 100MB”. Agents are at their best when they can investigate; over-specifying steps transfers your guesses into the plan.
Define done
Give an acceptance test: “done when the test suite passes and a 100MB export completes”. Without a finish line an agent either stops early or gilds the lily. This is the single highest-leverage line in the brief.
Set the boundaries
Say what is off-limits: “do not touch the billing module”, “no new dependencies”, “ask before deleting anything”. An agent with tools can do more damage than a chatbot with opinions, so the constraints section works like a permissions system written in English. Enterprises are learning this the expensive way, as we covered in OpenClaw’s $1.3 million agent bill and the agentic AI buyer’s guide.
Give it the context a new hire would need
Which framework, which conventions, where the tests live, what “good” looks like in this codebase. Paste error messages whole. The agent can read your repository, but it cannot read your intentions.
Related resources
- How to write a prompt that works: the general structure this builds on.
- Prompt templates: reusable starting points.
- What “agentic” actually means.