The One-Page Brief That Actually Ships With An AI Agent

Your AI agent is not underperforming because the model is bad. It is underperforming because your brief is vague.

We see the same pattern in every client engagement. The owner asks the agent to "handle customer support." The agent does something that looks like customer support. Half of it is wrong. The owner blames the agent. The real fault is upstream.

An agent will execute against whatever brief it can construct from the instructions you give it. If the brief is fuzzy, the execution is fuzzy. If the brief is precise, the execution has a chance.

Here is the one-page format we use to brief any agent that ships real work.

#Five fields

Acceptance criteria. Not "make it good." Not "handle it." A list of concrete conditions the output must meet. "Response is under 200 words. Response links to the pricing page. Response never quotes a price the CRM does not confirm." Three to seven bullets. Every bullet is checkable by a human in under thirty seconds.

Constraints. The things that cannot be traded away for other goals. Budget. Deadline. Legal restrictions. Brand voice rules. If an agent could improve output by ignoring one of these, list it here so it does not.

Input format. Exactly what the agent will receive. "A Gmail thread with at least one inbound message. Sender name in the From header. Attachments may or may not exist." If your agent gets input from multiple systems, list each source and its schema. Vagueness at the input is the single biggest cause of silent failures.

Output format. Exactly what the agent will produce. "A Gmail draft in the same thread. Subject prefixed with 'Re: '. Body under 200 words. No attachments. Signature block from templates/founder-sig.txt." The agent should not have to guess where the output goes or what shape it takes.

Escalation rule. When to stop and ask. "If the customer requests a refund over $500, do not send the draft. Post the thread to Slack channel #escalations and tag the CEO." Every agent needs a stop condition. Without one, agents keep trying to fulfill impossible briefs and produce silent garbage.

#Why this works

The brief format is boring on purpose. Boring is legible. An agent given a five-field brief has almost no room to invent the wrong interpretation. A human reading the brief can spot ambiguity before the agent starts, not after it ships.

You will spend twenty minutes writing your first brief. You will spend two minutes writing your tenth. The first ten make the next hundred cheap.

#What most owners skip

Two fields get skipped constantly.

The escalation rule gets skipped because "the agent should figure it out." It cannot. Every agent trained to complete tasks will complete a task, even a bad one, unless you explicitly tell it not to.

The output format gets skipped because "the agent knows what an email is." It does not, at the level of specificity that keeps your inbox out of trouble. Names of files, position of the signature, whether attachments are allowed. These matter and they will not be inferred correctly.

#The takeaway

If you are running an AI agent and the output is not what you want, do not switch models yet. Rewrite the brief on one page with the five fields above. In our experience, most "the agent is not smart enough" complaints turn out to be "the brief was not clear enough" once you look at the actual instructions.

If you want help auditing the briefs behind your current agents, or building the first one for a new agent, reach out at /get-started. One page. Five fields. That is usually where the fix starts.