A good agent brief has a finish line.

Name the outcome before the activity

“Research this” describes an activity. “Give me a one-page decision brief with three options and the trade-offs” describes a result. A result helps both you and the agent decide when the task is complete.

Use a format that fits the decision. A short table works for comparable options. A draft works for writing. A list of findings with references works for an investigation. Do not request extra output simply because the agent can produce it.

Use this six-part structure

1. Outcome: what should exist when the task is finished?
2. Sources: which documents, pages or accounts should the agent use?
3. Scope: which dates, projects or people are included?
4. Boundaries: what may it change, and what needs your approval?
5. Output: where should it put the result and in what format?
6. Verification: what evidence should it show so you can check the work?

Not every task needs six headings. For a small request, a few sentences can cover all six. The point is to remove ambiguity that could change the result.

An example you can adapt

Compare the three proposals I attached for our website refresh. Use only the current proposal versions. Summarize scope, timeline, assumptions and open questions in a table, then give me a short recommendation. Do not contact the vendors or accept anything. Link each finding to the relevant proposal and flag terms you could not verify.

The agent has a target, a source boundary and a stopping point. It can make progress without deciding whether “research” also means sending questions or committing to a supplier.

Match permissions to the task

A writing brief generally does not require full control of your Mac. Begin with the access needed to read the material and prepare the result. If the task later needs a protected action, review the target and effect before allowing it.

YankyBot offers read-only, ask-first and full-access modes. These settings and your written instructions serve different purposes: the mode controls tool access, while the brief describes the job. A prompt alone is not a substitute for an access setting.

Review uncertainty as well as output

A useful agent should say which source was unavailable, which number could not be confirmed and which assumption changed the answer. Ask for those gaps explicitly. A polished paragraph without evidence is not the same as a checked result.

After a successful run, keep this brief as a template. Put lasting context in editable memory and replace the task-specific inputs each time.