AI Creative Briefs: A Workflow for Ad Testing

What a testable brief contains

A testable brief answers six questions in plain words. If one field is vague, the test result will be vague too. An audience written as people who want to get fit gives a designer and a model almost nothing to work with. New parents with 20 minutes a day gives both a scene, a problem and a tone.

• Audience: one specific group and the situation they are in.
• Angle: the single reason this group should care.
• Hook: the first line or first three seconds that delivers the angle.
• Proof: the demonstration, feature or customer statement that backs the claim.
• Offer: the trial, discount or next step you are asking for.
• Metric: the one number that decides the result, such as cost per install or cost per trial.

Keep to one variable per test. Change the hook and hold the angle, format and offer fixed. If you change the hook and the offer together, you cannot tell which change moved the metric. Larger changes belong in a separate test with their own brief.

Write the decision rule before launch, while no result is pulling on your judgment. Illustrative example. Your target cost per install is $4 and each variant gets $300 of spend. The rule reads: kill a variant if its cost per install is above $6 (1.5 x $4) after $300, and scale it if it is at or below $4. Variants between $4 and $6 get a second budget step or a rewrite. The thresholds depend on your vertical and margins, so set your own and put them in writing before the first impression.

Inputs worth feeding the model

A model drafts from what you give it. Generic input produces generic angles, so collect three sources before you open the prompt.

• Top and bottom creatives with metrics: list winners and losers together with spend, the brief's metric, and a short description of hook, format and proof. Losers matter because a model that sees only winners will copy surface traits that may not have caused the result.
• Competitor angles from public libraries: browse the ad libraries the platforms publish, note the themes that repeat, and summarize them as angles. Use them as a signal of what the market is saying, and do not copy wording or visuals.
• Customer language from reviews and support: collect the phrases people use to describe the problem and the outcome. Strip names and personal data before pasting anything, and check your model provider's data policy.

Keep the input short and labelled. A table of 10 creatives and 20 customer phrases beats a pasted export of everything, and it costs fewer tokens.

The drafting workflow

Run the model in three steps, one prompt each, and review the output between steps. Step 1: ask it to summarize winners and losers, listing patterns in hook type, format and proof, and flagging where the sample is too small to trust. Step 2: ask it to propose four angles, each tied to an audience pain and to a specific line from your inputs. Step 3: ask it to write three hook variants per angle, each under a set length and each testing a different opening move.

Then a person edits. Check each claim against what the product does, adjust tone to your brand, and remove anything that conflicts with platform rules. The model produces candidates. Your team decides which ones deserve budget.

Illustrative example of sizing the first round. Four angles with three hooks each give 4 x 3 = 12 variants. Testing all 12 at $300 each costs 12 x $300 = $3,600. Say you start with six: the two angles your team rates highest, with all three hooks each, so 2 x 3 = 6 variants and 6 x $300 = $1,800. Hooks inside one angle compare cleanly. The comparison between the two angles is rougher, because each angle carries three hooks. The other six variants wait in the queue for the next cycle.

Quality checks

Run a short checklist on each draft before it reaches a designer or editor. The checks below take a few minutes and catch the problems that waste spend or put an account at risk.

• Claims: remove anything you cannot verify, including superlatives, rankings, before-and-after results and invented testimonials.
• Prohibited promises: cut promises of health outcomes or income, and any wording that presents a result as certain. Regulated verticals carry stricter rules, so read the policy for yours.
• Platform policy: compare each draft with the current ad policies of the platform where it will run. Passing your own check does not mean the platform will approve the ad, so plan for rejections.
• Brand voice: compare the draft with three approved ads and your list of banned words. Models drift toward a smooth, generic tone, and a human ear catches it fast.

Naming and tracking

Naming is what makes the test readable later. Use a name that encodes angle, hook, format and version, for example B014_A2_H3_UGC9x16_v1. That reads as brief 14, angle 2, hook 3, user-style video in 9x16, first version. Anyone filtering the ad account can see what a variant tested without opening it.

Link each variant to its brief ID in a sheet or database. The brief holds the hypothesis, the metric and the kill and scale rule, so the result can be judged against what you wrote before launch.

Report by angle, not only by ad. A single ad can win or lose because of delivery noise, while an angle with several variants shows a pattern. Sum spend and the brief's metric per angle, then read the ad-level rows as detail under that summary.

Closing the loop with a weekly review

Hold a weekly review that marks each variant as winner, loser or inconclusive against the rule in its brief. Winners go into the next brief as inputs, with the hook, format and proof described in a line. A creative team that runs testing cycles weekly builds a library of evidence this way, and each brief starts from more data than the last.

Archive losers with a reason: weak hook, wrong audience, policy rejection, too little spend, or poor execution. A reason turns a dead ad into a note you can search later and stops the team from retesting the same idea by accident.

Track hit rate per angle, meaning the share of its variants that met the scale rule. Say an angle produced six variants and two met the rule: 2 / 6 is one third. Small samples are directional, so treat the figure as a prompt for another round and as no verdict.

Next step: run one brief this week

Pick one live campaign and run the full workflow once before you build anything larger.

• Export your top five and bottom five creatives with spend and the metric you care about.
• Collect 20 customer phrases from reviews and support tickets, with personal data removed.
• Fill in the six brief fields and write the kill and scale rule with real thresholds.
• Draft four angles with three hooks each, edit them against the quality checklist, and launch six variants.
• Book the weekly review now, so the results feed the next brief.