← Research library

Writing briefs agents can actually follow

AI-assisted research. This note was prepared with the assistance of AI tools. It may contain errors and is not professional advice. See our Legal Disclosures.

The creative brief is one of marketing's oldest tools, and most teams already have a template for it. As agents take on drafting work, teams often hand them the same brief and are surprised by the results.

Agents can read a brief perfectly well. The trouble is that human briefs lean on shared context: the history of the account, the unwritten preferences of a stakeholder, the claim everyone knows not to make. Agents do not have that context unless you give it to them, so a brief that works for agents makes the implicit explicit.

Where human briefs fall short

A brief written for a colleague is a compressed document. It assumes the reader sat in the kickoff meeting, knows which customer stories are approved, and can tell whether "keep it punchy" means two paragraphs or a full page. Experience and quick questions across the desk fill the gaps.

An agent fills the same gaps by guessing. Its guesses are usually plausible, which is what makes them hard to catch. A draft can cite a customer result nobody approved, or adopt a tone the brand dropped last year. It can read well and still be wrong in ways a reviewer has to know to look for.

The fix is structural. Instead of writing longer briefs, write briefs with sections that each close one specific gap.

The agent-ready brief

We use a brief with seven sections. Each one is short, and each one answers a question an agent would otherwise guess at:

  1. Objective: one sentence describing the outcome, ideally with a metric, such as "Drive registrations for the June webinar from existing customers."
  2. Audience: who exactly, including what they already know and care about, rather than a bare label such as "decision makers."
  3. Message: the single most important takeaway, plus no more than three supporting points.
  4. Proof: approved facts, claims, and examples the agent may use, with anything unlisted treated as unverified.
  5. Constraints: length, format, channels, legal requirements, words to avoid, and anything off limits.
  6. Voice: a link to the brand voice specification, plus any adjustments for this piece.
  7. Definition of done: the checks the reviewer will run before approving.

The voice section should point to a shared document rather than restate style rules in every brief. A brand voice specification kept in one place can be updated once and used everywhere. The brief then only needs to record what is different about this piece.

The proof section does the most work

If you change one thing, make the proof section strict. Agents are good at producing plausible statements. Without a defined set of approved facts, they will fill gaps with confident claims about your product, your customers, or your results that nobody has verified.

A good proof section lists four kinds of material:

  • Product capabilities: what the product does, in plain language, with any limits stated.
  • Approved customer examples: each with permission noted and the wording that may be used.
  • Citable numbers: each with its source and the date it was last checked.
  • Prohibited claims: statements the agent must not make, even if they seem to follow from the facts.

Then add one instruction: if a needed fact is not in the proof section, flag it instead of inventing it. Flags are useful data. A brief that triggers many flags is missing material, and when the same facts go missing across briefs, the gap is in your shared knowledge base rather than in any one brief.

For recurring facts, link to a maintained source instead of pasting them in. Pasted facts go stale inside old briefs, while linked ones stay current.

A worked example

Consider a hypothetical software company briefing an agent on a launch email for a new reporting feature. The first version of the brief says: "Announce the new reporting feature to customers. Highlight the time savings. Keep it short and on brand."

An agent given that brief has to guess at most of the job. It does not know which customers to address, how much time is saved, whether any figure is approved, how short "short" is, or what the reviewer will reject. A likely result is a fluent email with an invented time-savings figure.

The agent-ready version sets the objective as feature adoption among admins on the current plan. It describes the audience as admins who already run weekly exports by hand.

It lists two approved capabilities, states that no time-savings figure is approved yet, and caps length at 150 words. Done means accurate claims, one call to action, and a working link. The agent now has little left to guess.

Briefs as reusable assets

Once briefs are structured, they become data. You can store them, compare them, and learn from them. Over a quarter, you can see which kinds of briefs produce drafts that need heavy editing and which produce drafts that ship nearly as written.

To make that feedback loop work, record a few things against each brief:

  • Edit rate: how much the reviewer changed the draft before approval.
  • Flags raised: how many missing facts the agent reported.
  • Rejection reasons: which definition-of-done check failed, if any.

Patterns in these records show where briefs need work. If rejections cluster on audience, the audience section is too vague. If edits cluster on tone, the voice specification may need examples. A starter rubric for content agents can turn the definition of done into consistent scores.

Briefs for humans benefit too

Briefs written for agents tend to be better briefs for people. New team members, freelancers, and agencies all benefit from explicit proof and a clear definition of done. Writing for an agent forces out ambiguity that was always costing time, and that cost was simply easier to absorb when everyone shared the same context.

The same structure also works between agents. When a research agent hands work to a writing agent, the handoff is effectively a brief, and the same sections apply. We discuss that contract in designing agent handoffs.

The question we are studying next is how much of a brief an agent can draft for itself. A planner agent could fill in the audience and constraints from past campaigns, leaving people to write the objective and approve the proof. We expect that split to work for routine pieces, and we are testing where it breaks down.

How to apply this: add a proof section and a definition of done to your existing brief template this week. Instruct agents to flag missing facts instead of filling them in, and track how often they do.

Related research

Put this research to work

Marketix turns these patterns into software: agents, automated workflows, and a hub that keeps people in control. Sign up for early access.