Writing Prompts Brief an AI Agent So It Can Finish Without You

Anyone delegating to AI agents · Hand off a task an agent can complete unsupervised

Brief an AI Agent So It Can Finish Without You

A checklist for the prompt you hand a coding or research agent, so it returns a result and not a promise to get to it.

You are writing the brief another AI agent will work from. It starts with no memory
of this conversation, so everything it needs to finish has to be on the page you
hand it. Every brief carries these seven things.

  Its role, in one line. "You are reviewing this change; you did not write it" is a
  different job from "you are writing it."

  Where it works: the project, the branch or folder, and what it may change. Name
  what it may not touch too.

  The decisions already made, dated, in the words of whoever made them. An agent
  that doesn't know a question is settled will reopen it.

  Earlier claims, each marked "verify, not fact". Anything a previous session
  concluded is a lead. Tell the agent which claims to check before building on them.

  What done looks like, as something a command or a read can show: "the test
  passes", "the page returns 200 and shows the new heading", "the report has one row
  per input file". If done needs someone's taste, say whose, and say the agent
  should stop there and ask.

  The check to run before it reports. Tell it to wait for the check to finish and
  read the output to the last line.

  What to report, and in what shape. One line of result, then the evidence. A failed
  or unrun check is reported as failed or not run, never as passing.

Write the brief for a capable stranger. Leave out your shorthand and the history of
how you got here, unless the history changes what it should do.

When the agent comes back with "I'll do that next" or "this should work", treat it
as a failed run, and run it again with the brief tightened at the point where it
went wrong.