The difference between a mediocre agent session and a reliable one is rarely the model. It is the brief. A good brief lets the agent plan, execute, and prove completion without guessing what you meant.
The five required parts
- Goal. One sentence describing the finished outcome, not the activity.
- Context. What exists already: repository, content, conventions, constraints.
- Requirements and out-of-scope. What must be true, and what the agent must not do.
- Acceptance criteria. Observable, yes/no checks that define done.
- Evidence. What the agent must show to prove each criterion passed.
If a brief lacks any of these five, the agent will invent it. Invented constraints are how projects drift.
Write criteria as checks, not feelings
“Make it look professional” is not an acceptance criterion. “No horizontal overflow at 390px” is. “Improve the content” is not. “Every factual claim has a source and a check date” is. When you review a brief, rewrite every vague word until each criterion is a check someone could run.
Scope the out-of-scope list
Agents expand scope by default. An explicit out-of-scope list is the cheapest guard: “no login system”, “no database”, “no new pages beyond the five listed”. List at least three things the agent should not do. The list makes it safe for the agent to say no.
Ask for evidence
End the brief with the evidence requirement: screenshots at named widths, the build log, the link check report, the checklist with pass and fail per item. Evidence turns the final review from a conversation into an audit. You are not asking the agent to be honest, you are making honesty checkable.
A template that works
Goal: [one sentence, finished outcome]
Context: [what exists, conventions, constraints]
Requirements:
- [observable requirement]
- [observable requirement]
Out of scope:
- [thing the agent must not do]
Acceptance criteria:
- [ ] [yes/no check]
- [ ] [yes/no check]
Evidence:
- [screenshots at 390 and 1440px]
- [build log from a clean checkout]
- [checklist with pass/fail per item]Fill in the brackets, hand it to the agent, and review the evidence, not the vibes. Every scenario page on this site includes a task brief generated from its own data, so you can copy one and adapt it instead of starting from scratch.