From Winning Ad to New Creative: An AI Swipe Workflow

Swipe the skeleton, not the skin

Every winning ad has two layers. The skin is the words, the faces, the footage and the specific claims. The skeleton is the order of the beats and the job each one does: hook, qualifier, credibility, problem, cause, solution, proof, offer.

The skin belongs to the advertiser, and most of it won't fit your product anyway. The skeleton is what made the ad work, and it transfers. A swipe workflow is a way of taking the skeleton cleanly and leaving the skin behind.

AI is well suited to that separation, because breaking a script into labeled beats is exactly the kind of reading it does well. The rest of this article is the workflow, step by step.

Step 1: choose a donor that has proven itself

A swipe is only as good as its donor. On our board (2,443 winning Meta ads, October 9, 2026, all with $50k+ in estimated spend), here's how the pool narrows as you add proof:

Days live and variants matter because they're signals advertisers can't fake: brands don't keep paying for an ad for three months, or keep cutting new versions of it, unless it works. Ads in the 90-day, 2-variant group reach our top tier 21.5% of the time against about 10% overall, though part of that is built in, since the winner score itself uses days and variants. We explain both signals in how to find winning Facebook ads.

Step 2: check fit before falling in love

A donor can be a great ad and a bad fit. Before mapping anything, answer four questions:

1. Awareness. Does the donor talk to the same awareness level as your audience? An unaware-audience story won't fit a product people already search for by name.
2. Proof. What does the donor use to be believed (a demo, a credential, a number, a customer)? Do you have the equivalent?
3. Price and offer. Is the donor selling a $29 impulse buy or a $300 considered purchase? Structures for one rarely work for the other.
4. Format. Can you produce it? A founder-expert ad needs a credible person on camera.

If two answers are "no", pick another donor.

Step 3: map the donor line by line

Short donors make the method easy to see. Here is a 27-second ad that has run almost a year.

The map, and how the same skeleton carries a different product (a slow-feeder bowl for cats):

Notice what transferred: the order and the jobs. Not one phrase. That table is the swipe.

Step 4: let AI do the annotation

For longer donors, mapping by hand is slow, and this is where AI saves the most time. Ask for the jobs, not a rewrite:

Here is a winning ad transcript. Do NOT rewrite it.
Split it into beats. For each beat give:
- the exact lines (quoted),
- its job: hook, qualifier, credibility, problem, cause, solution,
  proof, objection, offer, CTA,
- the copy block it serves: pain, promise, proof, constraint,
  curiosity,
- what the viewer sees, if you can tell.
Then write the skeleton as one line per beat, with no donor words.

If you use Claude with the Ad Radar MCP server, get_swipe_prompt does this setup for you. Given an ad and your product, it returns a prompt with the donor script annotated by copy block (pain, promise, proof, constraints, curiosity) and instructions to work in order: structure first, then a full draft, then a comparison with the original. The full research session around it is in competitor ad research inside Claude.

Step 5: rebuild for your product, across categories if you like

The skeleton is portable further than most people think. Here's a longer donor from home appliances:

The skeleton in one line per beat:

1. An expert stops recommending a common fix.
2. Qualify: "if [symptom] no matter what you've tried".
3. Credibility: who I am, how long, how many cases.
4. The pattern I keep seeing.
5. Why the common fix fails.
6. What actually causes it.
7. What I recommend now (the product).

Now carried into a different category, say a car-care product with a detailer on camera: "I stopped telling customers to take their cars through automatic washes about a year ago. If your paint looks dull a week after a wash, give me two minutes..." Same seven jobs, nothing borrowed but the order. Every fact in beats 3 to 6 has to be true for your expert and your product.

Prompt for this step:

Skeleton: [7 lines from above]. Product facts: [list]. Our expert:
[who, real credentials]. Write a [length]-second script that
follows the skeleton beat for beat. Rules: no phrase of 4+ words
from the donor; every claim from the facts list or [NEEDS FACT];
spoken English, short sentences.

For keeping the draft sounding like a person rather than a model, see writing video ad scripts with AI.

Step 6: run a distance check

Before anything goes to production, check the draft against the donor:

Compare my draft with the donor transcript.
1. List any run of 4+ identical words.
2. List any claim, number or name that came from the donor.
3. List visuals I described that copy the donor's shots.
4. Score overall similarity: structure only / structure + phrasing /
   too close. Explain in one line.

You want "structure only". Anything closer gets rewritten. Beyond the legal and ethical side, there's a practical reason: Meta's delivery system is understood to group ads that look and sound alike, so a near-copy of a competitor's ad competes against it rather than finding new people. The industry calls this a creative similarity score, though it's a practitioner term, not an official Meta metric. More in creative similarity score.

Step 7: change one big thing

The last step makes the swipe yours. Keep the skeleton and change one major element the donor didn't use:

• Hook pattern. Donor opens with a story? Open with a number.
• Production format. Donor is footage with voiceover? Shoot it as a founder to camera.
• Proof type. Donor tells? You show, on camera.

One big change plus the skeleton gives you an ad that is proven in structure and new in execution. It also makes the swipe easier to learn from: when it wins or loses against the donor's pattern, you know which variable moved. More on running that as a test in iterations vs new concepts.

The workflow on one page

The AI does the reading and the first draft. You choose the donor, supply the truth about your product, and decide what makes it different.

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.