How to Write a Creative Brief with AI (from a Winning Ad)

Start from an ad, not from a blank page

Ask a model to "write a creative brief for my protein drink" and you get a brief that could describe any product: aspirational lifestyle, authentic UGC, strong hook, clear CTA. Every field is filled and none of it is a decision.

The fix is to make the model reverse-engineer one ad that already works, then reuse its decisions for your product. A winning ad is a set of choices someone already tested with real money: who it talks to, how it opens, what it proves, where it sends the click. A brief is the same set of choices, written down in advance.

This article is the prompt chain we use for that. If you want the blank template itself, it's in our creative brief template. Here we focus on getting a model to fill it well.

Which fields deserve the model's attention

Not every brief field matters equally. We grouped the 2,443 ads on our board by each tag dimension and compared how often each choice reaches the top tier (top 10% by winner score, live 21+ days). The spread between the best and worst common option (40+ ads) shows how much a field can move the outcome.

Two caveats. Production format is tagged on only 610 ads, so treat that row as directional. And these are correlations across brands, not causes. Still, the order is useful: when the model fills format and proof, make it justify the choice with evidence from the donor. When it fills awareness, a sensible default is fine.

The four-prompt chain

We split the work into four prompts instead of one. One giant prompt produces a brief that looks complete; four short ones produce a brief where you can check each step.

Prompt 1: extract the decisions

Paste the donor ad (script, headline, body, estimated spend, days live, variants) and ask:

Here is a winning Meta ad. Do not rewrite it. Extract the decisions
behind it, one line each, quoting the script where possible:

- Audience and awareness level (unaware / problem / solution /
  product / most aware). What does the opener assume they know?
- Narrative angle (root cause, systematic breakdown, story, expert,
  wrong question, cautionary tale, insider, accidental discovery).
- Hook pattern and the exact opening words.
- What the viewer SEES in the first 2 seconds.
- Body beats in order, with the job of each beat.
- Proof: what makes the claim believable, and is it shown or told?
- Offer and CTA. Where does the click go?
- Production format and length.
If something is not in the material, write "not visible".

The last line is the most important one. Models fill gaps. "Not visible" is the answer that keeps the brief honest.

Here is the donor we used for the worked example below.

Run Prompt 1 on it and the decisions come out cleanly: solution-aware viewer (already buys refreshers), a race against a named alternative as the angle, a question opener, the shaker and the drive-through order visible from the start, body beats that follow the recipe (water, scoop, shake), and proof by demonstration plus a nutrition line ("10 grams of protein, fiber, electrolytes, and zero grams of sugar").

Prompt 2: map the decisions to your product

Our product: [facts, price, offer, what it is NOT allowed to claim].
Audience: [who]. Test question: [one sentence].

For each decision you extracted, say:
KEEP (works as is for our product), ADAPT (how), or DROP (why).
Any proof we don't have must be marked MISSING, not invented.

This is where most of the value is. The model has to defend each choice against your product's real facts. A race-the-coffee-shop angle only works if your product is genuinely quick to make. If it takes four minutes, the model should mark the angle DROP, and you've just saved a shoot.

Prompt 3: fill the brief

Fill this brief template with the KEEP and ADAPT decisions.
[paste template]
Rules:
- Three openers in three DIFFERENT hook patterns, exact words.
- The first-frame line describes something a camera can film.
- Proof field lists only proof we have. Put MISSING items under
  "Needed before shoot".
- One page. No adjectives in the angle field.

"Three different hook patterns" stops the model from writing one opener three ways. "Something a camera can film" stops it from writing "an engaging visual". For hook patterns that do well on the board, see Facebook ad hooks.

Prompt 4: attack your own brief

You are the media buyer who will judge this ad. List the 5 weakest
points of this brief: anything vague, unfilmable, unproven, or
likely to get the ad rejected. Suggest a fix for each in one line.

The critique pass catches what the first three missed, and it is fast. Fix, then hand it off.

A filled example

Here is the brief Prompt 3 produces when the product is a powdered matcha latte mix with a first-order discount, and the donor is the ad above. Fields shortened for space.

TEST QUESTION   Does a "faster than the café" race beat our current
                ingredient-led video for cold traffic?
AUDIENCE        Women 25-45 who buy iced matcha 3+ times a week.
                Solution-aware: they know the drink, not our mix.
ANGLE           Race against the café order. Same drink, less time,
                less money.
OPENERS         A (question): "Can I make an iced matcha latte before
                my order's ready?"
                B (number): "$6.75 at the café. 41 cents at home."
                C (story): "I timed my café run last Tuesday. Then I
                timed this."
FIRST FRAME     Phone timer starting next to a café cup and our tin.
BODY            1 Order placed (show screen). 2 Scoop, milk, shake.
                3 Taste test side by side. 4 Timer stops. 5 Price.
PROOF           Timer on screen (real). Price per serving (pending).
NEEDED BEFORE   Price per serving signed off by finance.
SHOOT
FORMAT          Creator selfie + overhead demo, 45-60s.
LANDING         Product page, first-order offer pre-applied.
KILL RULE       CPA > 1.3x target after 2x target CPA spent.

Notice what the model was not allowed to do: invent a nutrition claim, invent a review count, or pick a format without a reason. The price line in opener B is a placeholder until finance confirms it, and the brief says so.

When the donor's proof is the hard part

The donor in our example proves its point by doing it on camera. Many top ads do. Strong proof reaches the top tier 16.3% of the time on our board, against about 10% overall, and shown proof is easier to brief than claimed proof.

When the extracted proof is something you don't have (a study, a celebrity, ten thousand reviews), the honest brief has a "Needed before shoot" line, or a different donor.

Iteration briefs: let the model do less

If the brief is an iteration of your own winning ad, the body is proven, and the model's job shrinks to new openers and new first frames. Give it the winning script and ask for openers in hook patterns the original doesn't use.

More on when to iterate and when to brief a new concept in iterations vs new concepts.

Doing it with Claude and live ad data

The chain above works with pasted ads in any model. If you use Claude, an MCP connection to an ad database removes the pasting. With the Ad Radar connector, Claude can find donors with search_ads, open the full transcript with get_ad and adapt the structure with get_swipe_prompt, which returns a ready swipe prompt for your product. The whole sequence is shown in competitor ad research inside Claude.

Checklist before the brief leaves your desk

• The donor ad is named, with estimated spend, days live and variants.
• Every brief decision traces back to a line in the donor or a product fact.
• Three openers, three hook patterns, exact words.
• The first frame is something a camera can film.
• Proof lists only what you have. Missing proof is flagged, not written around.
• Production format and proof level have a stated reason (they move results most on the board).
• One test question and one kill rule.

Figures marked as estimated spend come from Ad Radar's model of engagement on public Meta Ad Library ads. They are estimates, labeled as such, and are best used to rank ads against each other.