The one-page agent brief: the document that decides whether your AI helps or flails

When an AI agent gives a customer a wrong answer, the model is rarely the problem. The problem is that nobody ever wrote the right answer down. Your team carries it in their heads — the pricing exceptions, the questions that need a human, the promises you never make. An agent cannot read minds, so it improvises. The fix costs one page and about forty minutes.

1. What we do, and for whom. Three sentences: your services, your customers, your service area and hours.
2. Ten common questions with approved answers. Take the ones that arrive every week and write the reply you would be happy to see sent.
3. Voice and languages. Formal or friendly, which languages you answer in, what you call your own products.
4. Never do this. No prices for custom work, no promised dates, no comparisons with competitors — be blunt, not diplomatic.
5. When to fetch a human. Name the triggers: complaints, refunds, anything above a set amount.

Paste that page into your assistant, ask it ten real customer questions, and fix whatever it gets wrong — in the brief, not in the AI. That loop is the whole job: my Telegram intake bot and the chat on this site both run on nothing more than a page like this.