How to brief AI agents: a brief template for conductors

To brief AI agents well, give them a purpose rather than a task list, define the boundaries they must not cross, describe the audience in real language, specify what good looks like and what to avoid, and set checkpoints where they must stop for a human. A strong agent brief reads like a good creative brief with explicit rules attached.

Why do most agent briefs fail?

Because they are written like prompts. "Write ten Instagram captions for our new protein bar" is an instruction. It gives an agent no purpose, no audience, no boundaries and no taste to aim for. The agent fills the gaps with averages, and you get average work back.

Agents also take briefs more literally than people. A human copywriter will ask what you meant. An agent will usually just proceed. Every ambiguity in a brief becomes a decision the agent makes for you.

What should an agent brief contain?

Items five and six do the most for quality. Agents respond strongly to examples, and naming clichés explicitly helps them steer away from the sea of sameness.

How do you describe an audience to an agent?

Use evidence, not demographics. "Women 25 to 34" tells an agent very little. "People who want to cook healthily on weeknights but feel guilty when they order in, and who describe themselves as 'too tired to think about dinner'" tells it a great deal.

That is why our briefs start from research. Our agents run on SOMIN, an AI audience-research platform, which surfaces tensions with the real posts behind them. SOMIN's CRM and audience tools help keep those descriptions consistent across loops. We explain the research stage in how AI agents do audience research without guessing.

What does a worked example look like?

A condensed brief for a creative agent:

Purpose: increase trial of a meal-kit service among busy parents; judged by first-order profit, not clicks. Audience tension: wanting to cook real food for the family while having no energy to plan it. Phrases from research: "decision fatigue", "same three dinners". Boundaries: no health claims; no guilt about takeaway; no implying parents are failing. Good: warm, specific, slightly funny. Avoid: "dinner, sorted", stock smiling-family imagery, countdown urgency. Volume: 30 headline and opening-line pairs. Checkpoint: stop after drafts for human edit. Owner: account conductor.

Notice the boundary "no guilt about takeaway". Guilt would probably perform. It is also exactly the kind of efficiency that costs humanity. Writing it into the brief means the agent never needs to be corrected for it.

How detailed should brand boundaries be?

Detailed enough to be checked. "Stay on-brand" cannot be enforced. "Never compare directly with named competitors" can. If your brand has a voice system, convert its principles into explicit rules for agents. Our sister consultancy GPT3 Marketing builds brand voice systems designed to be usable by both humans and AI, and their approach is worth a look.

How often should briefs be updated?

Whenever the human edit keeps correcting the same thing. If a conductor rejects variants for the same reason three weeks running, that reason belongs in the brief. Briefs are living documents in an agentic setup; the edit loop feeds the brief loop. We cover the editing side in why AI creative needs a human edit.

Where can I learn more?

For agencies, SOMIN for agencies shows how audience research flows into briefs, and the Blak Labs case study is a good example of a creative agency using AI insight. The Marketing Mondays community also decodes campaigns weekly, which is useful raw material for the "what good looks like" section of any brief.

What are the most common briefing mistakes?

• Briefing the task, not the outcome. "Write captions" instead of "increase trial among people who feel X".
• Describing the audience by demographics alone. Agents need tensions and language, not age bands.
• Leaving boundaries implied. Anything you would be upset to see should be written down.
• Asking for too much volume. More options than a human can genuinely review turns the edit into skimming.
• No owner. A brief without a named conductor has nobody to update it when the agent drifts.

Most of these mistakes come from treating agents as fast copywriters rather than as collaborators who need context. Brief them the way you would brief a talented new colleague on their first day, then add the rules a talented new colleague would know without asking. The extra ten minutes at the start saves hours of correction later.

A final tip: test every new brief on a small batch first. Ask the agent for five options, read them, and fix the brief before asking for thirty. Ten minutes of calibration catches most misunderstandings, from tone that is slightly off to a boundary the agent read too narrowly, and it means the full run arrives closer to what the conductor actually wanted.

When you are ready, brief the agents at ask@agentc.marketing. A human conductor reads every brief first.