Scalewell / Blog

AI creative brief generator: how to turn ad performance data into briefs

A field-by-field brief template, the prompt that drafts it from your tagged ad performance data, a labeled worked example, and the five things the brief still needs from a person.

Data points flowing from a laptop into a single sheet of paper that a hand is editing, representing performance data turning into a creative brief

An AI creative brief generator writes the brief for your next round of ads from a structured set of inputs. The version worth using starts from your own ad-level performance data, tagged by creative element and ranked, and produces one hypothesis-led brief per concept. Fed only a goal and an audience, it writes the same brief any competitor could generate.

You know the Monday this happens. The weekly report shows cost per lead on Meta up a third, the designer asks what to make next, and what she gets is a screenshot of the three ads that held up and a note saying "more like these". That note is the brief, and because nobody wrote down why those three worked, the next batch is a guess dressed as a plan.

This article is for the head of paid media who wants a repeatable way to turn last quarter's numbers into this quarter's creative. It gives you the brief template in full, the data to pull from Meta, Google and LinkedIn, the prompt that writes the first draft, a worked example, and a plain list of what the brief still needs from a human. With the exports, tagging pass and brief prompt in a monthly rhythm, the team starts each round of creative from evidence, and Scalewell helps put that rhythm in place with whatever tools it uses.

Most generators brief from a goal and an audience, which is why the output is generic

A creative brief generator that asks for goals, audience and deliverables will return objectives, key messages and deliverables, because that is all it was given. ClickUp's generator, for example, takes "client goals, target audience, deliverables, and brand guidelines" and produces "a structured draft covering objectives, key messages, and deliverables", according to its own feature page. For paid media that misses the one input that matters, which is what your last fifty ads did.

The drafting step itself has become a commodity. A July 2026 piece from Digital Applied argues that Asana, Jasper, HubSpot, Monday, Wrike and Adobe all ship brief drafting now, so the difference lies in the inputs you supply and the review you enforce. Their summary is four words long: "AI drafts. Senior judgment approves." The rest of this article is about making the inputs good enough that the second half has something worth approving.

Isaac Rudansky of AdVenture Media puts the usual failure plainly in his guide to Meta briefs: "A creative brief fails when it describes what the brand wants to say rather than what the audience needs to hear." Your performance data is the closest thing you have to the audience telling you what it heard.

Four things to have in place before you generate anything

An ad-level export with the creative attached, a short tagging scheme, a minimum-data rule and a fixed template are the four inputs the generator needs, and the template is the only one most teams already have.

An ad-level export with the creative visible. Every row is one ad, and each row carries the copy and a link to the visual. All three ads managers export ad-level rows; add the copy and the creative link yourself if the export leaves them out, and step 1 below says what to pull from each platform. An API is only worth it once the method is running monthly.

A tagging scheme of six to ten elements. Hook type, format, whether a person is on screen, the kind of proof used, the offer, and the call to action cover most B2B accounts. The AI does the first pass from the copy and thumbnails, and a human checks a sample. Tags turn a list of ads into averages per element, which is what the brief is built on.

A minimum-data rule. Meta's learning phase documentation says ad sets usually leave learning "after about 50 results in the week after the ad set's last significant edit", and that performance during learning "is less stable". Ads below your floor go in a separate pile labeled "not enough data", never into the averages. Pick the floor per metric before you look at the numbers, not after.

The template. The next section gives it in full.

The brief template: one concept per brief, thirteen fields

Each brief covers one concept and states a hypothesis with a metric and a threshold, so the next round of data can prove it wrong. The third column says who fills each field, because that is where most AI brief workflows blur.

FieldWhat goes in itFilled by
1. Concept nameA short name you can find in the ad name laterHuman
2. Performance contextWhat ran in the period, which elements won and lost, with the numbers and the date rangeData, written up by AI
3. Hypothesis"If we [change], then [metric] will [move by X] against [control], because [reason]"AI drafts, human edits the reason
4. Audience and stageWho sees it, how aware they are of the problem and of youHuman
5. Core messageOne claim, in one sentenceHuman
6. Hook optionsThree opening lines, each built on a winning hook typeAI drafts from winners
7. KeepElements the data says must stay, with the figure next to eachData
8. AvoidElements the data says to drop, with the figure next to eachData
9. Format and specsPlacement, size, length, text limitsHuman or template
10. Offer and call to actionWhat the reader gets and the exact button textHuman
11. Landing pageURL and the one thing the page must say in its first screenHuman
12. Test designWhat varies, what stays constant, minimum spend before judging, decision dateAI drafts, human approves
13. Owner and due dateWho signs it off and when creative is dueHuman

The data fills three of the thirteen fields, the AI drafts three more, and the human fills the remaining seven, which should take about twenty minutes.

A one-page brief card divided into thirteen color-coded bands, with three hands filling different parts of it
The thirteen fields by who fills them. Data fills three: performance context (2), keep (7) and avoid (8). AI drafts three: hypothesis (3), hook options (6) and test design (12). A person fills the remaining seven: concept name (1), audience and stage (4), core message (5), format and specs (9), offer and call to action (10), landing page (11), owner and due date (13).

The seven steps run from export to a signed brief in an afternoon

In our estimate the whole sequence takes an afternoon the first time and under an hour once the prompt and tags are saved.

  1. Pull 30 to 90 days of ad-level data. For Meta, request spend, impressions, reach, clicks, ctr, cpc, cpm, frequency, conversions and cost_per_conversion at ad level; these are the common fields Meta lists in its Ads CLI insights documentation, and any valid Insights API field can be added. If you run dynamic creative, add the asset breakdowns (image_asset, video_asset, title_asset, body_asset, call_to_action_asset) listed in Meta's asset-level insights documentation. One caveat when comparing old and new ads: Meta announced in October 2025 that from January 12, 2026 the 7d_view and 28d_view attribution windows are no longer returned, so set one window for the whole comparison. For Google, open each responsive search ad (up to 15 headlines and 4 descriptions) and select "View asset details" to reach the ad-level asset report, which now gives full statistics per asset in place of the old "Performance label" column. For LinkedIn, call adAnalytics with pivot=CREATIVE, which needs the r_ads_reporting permission and takes up to 20 named metrics, per the reporting documentation.
  2. Attach the creative to every row. Primary text, headline, description, a thumbnail link and the landing page URL, because a row without its copy cannot be tagged.
  3. Apply the floor and park the learners. Ads under your impression or result floor go in the "not enough data" pile. On Meta, add the "Last significant edit" column in Ads Manager and park anything still in learning.
  4. Tag the elements. Run the AI over the copy and thumbnails with your six to ten tags, using the tagging prompt that follows this list, then check one row in ten yourself and fix the tagging rules rather than the individual rows.
  5. Rank by the metric closest to pipeline that exists at ad level. Cost per qualified lead beats cost per lead, which beats click-through rate. Then compute the average of that metric for each tag value, with the number of ads behind each average.
  6. Run the prompt. It is below. Paste the tag table and the top and bottom five ads with their copy.
  7. Hold the review gate. A senior person edits the hypothesis, rewrites the "because", fills the human fields and dates the brief. Digital Applied's observation that "the review gate is the step teams skip" is the right warning.

Step 4 only works when every tag is defined in writing, with an example of what it is and what it is not, before the model sees an ad. Rudansky's brief defines the hook axis as "Opening 3 seconds of video or first line of static copy", and that is the unit you tag. The pair most likely to blur is statistic and bold claim. As of November 2026 Motion's hook library, now published by Runneth, describes a bold claim as one that "makes an outsized, extreme, or superlative promise" and a statistic hook as one that uses "quantified evidence". Our rule follows from them: a statistic hook opens with a measured figure and says whose figure it is, as in the invented line "Acme closed its books in 3 days instead of 9 last quarter", and a bold claim opens with a promise nobody is shown to have measured, such as "Cut your close by 80%".

Anthropic's prompting guidance says a few well-crafted examples "improve accuracy and consistency" and suggests three to five, so the prompt ends with hand-tagged ads. Plan to rewrite the definitions once, too: in a 2024 study of LLM annotation revised in 2026, Cheng, Mayya and Sedoc found model agreement with human labels was lower under first-draft guidelines than under guidelines refined over rounds, by 0.1 to 0.5 on Cohen's kappa. The prompt below works the hook element in full, with invented example lines, and gives the shape to copy for the other elements.

You are tagging B2B ads so that performance can be averaged by creative element.
Tag one ad at a time from its primary text, headline, description and the
thumbnail or a written description of it. Return one line per ad: ad_id, one
value per element, then a confidence of high or low. Use only the values defined
below. If no value fits, write "other" and quote the words that made you hesitate.

ELEMENT 1: hook_type. The hook is the first line of the primary text, or the
first 3 seconds of a video. Tag the first line only, whatever follows it.
- statistic: opens with a measured figure and says whose figure it is or where
  it was measured. The test is that someone measured it; a number alone does
  not qualify. A customer count is social_proof, not statistic.
  Example: "Acme closed its books in 3 days instead of 9 last quarter."
  Also statistic: "72% of finance teams still reconcile in spreadsheets (2026 survey)."
- bold_claim: opens with a promise, a superlative or a flat assertion that
  nobody is shown to have measured, with or without a number.
  Example: "The fastest month-end close in finance."
  Also bold_claim: "Cut your close by 80%."
- question: opens with a question to the reader.
  Example: "Still closing the books by hand?"
- pain_point: opens by naming the reader's problem in their words, as a statement.
  Example: "Month-end close still eats the first week of every month."
- social_proof: opens with a customer name, quote, count or award.
  Example: "Trusted by 400 finance teams."
  A measured result attributed to a named customer counts as statistic.

ELEMENT 2: {next element, in the same shape: name, a one-sentence definition,
the test for the boundary case, one example of what it is and one of what it is not}

(repeat for each of your six to ten elements)

HAND-TAGGED EXAMPLES
{three to five ads your team tagged by hand, each with its values and one line on why}

ADS TO TAG
{paste}

The brief prompt below is written for any capable model, and it asks the model to separate what the data shows from what it is inferring, which is the habit that keeps AI briefs honest.

You are helping a B2B paid media team write a creative brief from performance data.

INPUTS
1. Period and channel: {channel}, {start date} to {end date}.
2. Primary metric: {metric, e.g. cost per qualified lead}. Lower is better.
3. Floor: ads with fewer than {N} impressions or {M} results are excluded and listed separately.
4. Tag table: for each tag value, the average primary metric, the average CTR,
   the number of ads and the total spend. (pasted below)
5. Top five and bottom five ads by primary metric, each with primary text,
   headline, format, tags and the figures. (pasted below)
6. Brand constraints and claims we cannot make: {list}.

TASKS
A. Performance context (120 words max). State what ran, which tag values won and lost,
   with the figures. Mark every sentence as OBSERVED (directly in the table) or
   INFERRED (your reading). Do not infer a cause for any tag with fewer than 3 ads.
B. One hypothesis in this form: "If we [change one element], then [primary metric]
   will improve by at least [X percent] against [named control ad], because [reason]."
   X must be justified by the gap in the tag table.
C. Three hook options, each built on a winning hook type, each under 12 words,
   each marked with the tag it comes from.
D. Keep list and Avoid list: element, figure, number of ads behind it.
E. Test design: the one axis that varies, what stays constant, minimum spend per
   variation before judging (use {spend floor}), decision date.
F. Open questions for the human: anything you needed and did not have, including
   the audience insight, the offer and whether the primary metric reflects pipeline.

RULES
Use only the figures provided. Quote no external benchmarks.
If the data does not support a hypothesis, say so and stop.
Write in plain sentences. No exclamation marks.

TAG TABLE
{paste}

TOP AND BOTTOM ADS
{paste}
Relay race track with runners passing a baton and the track looping back to the start, representing hand-offs between platform, AI, spreadsheet and senior reviewer
The seven steps and their hand-offs: 1 export ad-level data (platform), 2 attach the creative to every row, 3 apply the floor and park the learners, 4 tag the elements (AI, with a human sample check), 5 rank by the metric closest to pipeline (spreadsheet), 6 run the prompt (AI draft), 7 hold the review gate (senior human). The test result is logged back into the next export, which closes the loop.

A worked example with invented numbers shows where the human changes the brief

Imagine a B2B payments company, call it Company A, spending $60,000 a quarter across LinkedIn and Meta on 24 ads, with cost per qualified lead as the primary metric and a floor of 3,000 impressions and 8 qualified leads per ad. Every figure in this section is made up to show the method. After the floor, 17 ads qualify. The tag table for hook type comes back like this:

Hook typeAdsAvg CTRAvg cost per qualified lead
Statistic from own data51.3%$140
Question60.9%$210
Bold claim40.7%$260
Customer quote21.1%$165

And for what is on screen:

On screenAdsAvg CTRAvg cost per qualified lead
Product interface90.8%$230
Named customer, face visible51.2%$150
Abstract graphic30.9%$205

The model's observed statements write themselves: statistic hooks beat question hooks by $70 per qualified lead across 11 ads, and a named customer on screen beats the product interface by $80 across 14 ads. Because the customer-quote row has two ads, it gets no inference under the rules. Its proposed hypothesis reads: "If we open with a statistic from the customer's own results and show the named customer rather than the interface, cost per qualified lead will fall by at least 25 percent against ad 14 (question hook, interface, $215), because both winning elements appear in the top three ads and neither appears in the bottom five."

Then the human reads it and asks the question the model cannot: why did statistic hooks win? Company A's head of paid media knows that the quarter's statistic ads all quoted reconciliation hours saved, a number finance buyers feel every month, while the question ads asked about "payment complexity", which nobody feels. That reason goes into the "because" and changes the hook options from "any statistic" to "a monthly-cost statistic the buyer already tracks".

The brief still needs five things only a person can supply

Performance data tells the brief what happened, and a person has to supply why it happened, what the offer is, what is new, what you are allowed to say, and whether the metric is even the right one.

The reason. Tag averages show that faces beat interfaces, and they do not know that your buyers are finance leads who have been burned by a tool that looked good and reconciled nothing. Field 3's "because" comes from someone who has sat in sales calls.

The offer. No export contains next quarter's offer or the pricing change, and the AI will invent one if the field is blank.

What is new. A generator working from your data can only recombine elements you have already run; if every ad for a year has used a product interface, the table has no row for "no interface at all". Someone has to add the untested idea on purpose, as its own concept, with a smaller budget and the same floor.

Claims, brand and legal. The model does not know your compliance team banned the word "guaranteed" in March, so put the banned list in the prompt and keep the final read with a person.

Whether the metric reflects pipeline. Cost per lead is the most common primary metric at ad level and the most misleading. Two ads with almost the same cost per lead can sit far apart once their leads are matched to opportunities in the CRM, and a brief ranked on cost per lead will call for more of whichever one produced cheap leads. If your CRM can give you cost per opportunity by ad, rank on that; if it cannot, say so in field 2.

Where data-led briefs go wrong

The failures are small samples, fatigue misread as a weak concept, learning-phase noise, and a winner briefed into the ground.

Too few ads behind an average. Two ads with a customer quote at $165 is a coincidence until it is six. The prompt's "no inference under three ads" rule is a floor, and the right number for your account depends on volume. If most of your tag values have two ads, tag less finely rather than briefing anyway; the OBSERVED and INFERRED labels in the prompt show you which sentences to delete.

A tired winner read as a bad idea. Meta labels an ad set or ad "Creative limited" when "cost per result is more than ads that you've ran in the past, but less than twice as much", and "Creative fatigue" when it reaches "twice as much", according to Meta's creative fatigue page. The same page says the label is only available for ad sets with one creative and that Meta counts "all recent exposures of the ad's image or video, including those from other campaigns from your Page". So an ad whose cost per result has reached twice that of your earlier ads after eight weeks at frequency 6 may be a winning concept that needs a refresh, and the brief should say "same concept, new execution" rather than "drop this". Check frequency and the date the decline began before the tag table gets to vote.

Learning-phase noise in the comparison. Meta lists "Any change to ad creative" and "Adding a new ad to your ad set" among the significant edits that restart learning, so every new brief produces ads that spend their first week in a less stable state. Judge them after the minimum spend in field 12, never in week one, and compare against a control that has run the same number of days where you can.

Briefing the winner into the ground. When a concept wins, the lazy brief is "ten more like it", and three rounds later the whole account is one idea at frequency 9. Meta's own fatigue recommendation is "a new image or video that is materially different from the original creative", so keep one concept slot in every round for something the table has never seen. And when you do iterate, change one axis at a time, as Rudansky advises: "In any single test, you should vary one axis while holding the others constant."

A brief worked when its hypothesis was tested and the result logged

A brief worked when the hypothesis in field 3 was tested at the spend floor in field 12, the primary metric moved against the named control by at least the threshold, and the result was written back into the brief.

That last clause is the one that compounds, because a brief with the result logged becomes a row in next quarter's table and the "because" lines accumulate into something like institutional memory. Rudansky suggests four to six test cycles to build a usable model of what works in an account; that is his view from agency work rather than a measured figure, and we'd expect the second and third briefs to be where the method starts paying for the setup.

If the hypotheses keep confirming, the thresholds are too soft; if the reviewer keeps rewriting field 2, the tagging is wrong and needs fixing upstream.

Set it up once as a project or a skill

As of November 2026, the template, the prompt and the tagging rules belong in a saved workspace so nobody pastes them from a document each week. In Claude, a Project holds project knowledge (your template, tag definitions and banned claims) plus custom instructions, and Anthropic's help center says projects are available on every plan, with up to five on the free tier. A custom Skill packages the same instructions in a SKILL.md file the model loads when the task matches; Anthropic's Agent Skills documentation says custom Skills can be uploaded in claude.ai with code execution enabled, created in Claude Code, or uploaded through the API. Our guide to Claude skills for paid media teams covers what to build first, and the prompt library for Google Ads and Meta Ads has the companion prompts for tagging.

The export can be automated too: Meta publishes an Ads CLI and an Ads MCP server whose reporting tools return campaign, ad set and ad level metrics including spend, impressions, CTR, CPC, CPM and conversions, per Meta's developer documentation. The routes are compared in Meta Ads MCP: connecting an AI client to your Meta ad account, and Mapping your paid team's workflows for AI shows where briefing sits among the other weekly jobs.

Where to start

Pull one platform's ad-level export for the last 90 days this week, tag it with six elements, and run the prompt once; the first brief will show you where your tagging and your floor need adjusting before the method is worth saving. Scalewell sets this up for paid teams, from the export and tagging rules to the saved project or skill the team runs each round.

Frequently asked questions

Can an AI creative brief generator work without performance data?

It can write a brief, and for a launch with no history it has to. Treat that brief as a set of hypotheses with no evidence behind them, size the first test to produce data rather than to win, and switch to the data-led method as soon as a dozen ads clear your floor.

How many ads do I need before briefing from data?

Enough that each tag value you want to act on has at least three ads above your impression and result floor. Six tag values at three ads each is about 18 qualifying ads, which most B2B accounts reach over 60 to 90 days. If you have fewer, use fewer tags, or extend the period before you tag more finely.

Should the brief be one page or several?

One concept per brief, and the thirteen fields fit on a page once the performance context is trimmed to its figures. Rudansky's fully populated template runs to four to six pages because it includes a hook library and iteration protocol; keep those as shared reference documents rather than repeating them in every brief.

Can AI tag creative elements reliably enough to trust the averages?

For text elements such as hook type, offer and call to action, a first pass from a capable model is usually consistent when the tag definitions are written down with examples. Visual tags such as "person on screen" need a human sample check. Fix disagreements by tightening the definition, and re-run the pass rather than hand-editing rows.

How often should I regenerate briefs from the data?

Once per creative round, and no more often than your floor allows. If a round needs two weeks to clear the spend floor, a weekly brief just reports noise. We'd run it monthly for most B2B accounts under $100,000 a month, and move to every two weeks only once more ads clear the floor each round.