Guide · 7 min read

Writing the perfect brief

Plain-language instructions that make an AI worker sound like your best team member, without drifting off-brand or off-policy.

A brief is the job description your AI employee reads before its first shift. The clearer the brief, the less time you spend correcting mistakes later. Good briefs combine facts, voice, guardrails, and examples. Bad briefs are a list of features copied from a help center.

Write it like a job posting

Start with the role, not the product. What is this AI employee hired to do? Who does it help? What does success look like? A brief for a sales AI should say: 'You qualify inbound leads, answer pricing questions, and book demos for the sales team. Success is meetings booked with qualified prospects.'

Include real sample conversations

Examples are more powerful than rules. Show the AI what a great reply looks like, what an acceptable reply looks like, and what a reply that needs escalation looks like. Five strong examples usually beat twenty abstract instructions.

  • A perfect answer to the most common question.
  • A polite but firm refusal of an out-of-scope request.
  • A handoff message that gives the human full context.
  • A correction from a real conversation that went wrong.

Define what it must not do

Guardrails prevent the most expensive mistakes. Be explicit about topics the AI should never touch, promises it cannot make, and data it cannot share. If a discount requires manager approval, say so. If a refund policy has exceptions, list them or escalate.

Set escalation as a feature, not a failure

The best AI employees know when they are out of their depth. Tell the AI exactly when to escalate: unknown pricing, angry customers, legal questions, anything involving money movement, or any request that contradicts policy. Escalation is not failure. It is the safety net that keeps customers happy.

A brief is a living document. The first version gets you to launch. The tenth version gets you to trust.

Iterate like a manager

After launch, read real conversations every day for the first week. Turn every mistake into a brief update. Did the AI offer a discount it should not have? Add a rule. Did it sound too robotic? Add a tone example. Within a month, the brief will be sharper than any onboarding manual you have written for a human.