Every operator I respect keeps a swipe file. A folder of ads that worked: screenshots from competitors, from categories three aisles over, from brands with a hundred times the budget. I kept one at every brand I've run. Most marketing teams treat it as inspiration, which is a polite word for a folder nobody has a method for.
The problem shows up the day you try to use it. Recreate a saved ad too literally and you've made a copy: someone else's joke, someone else's claim, someone else's brand wearing your logo. Recreate it too loosely and you lose the thing that made you save it in the first place. Teams bounce between those two failures for years without naming either one, because nobody ever told them what a winning ad actually contains.
A winning ad carries two layers
Look hard at any ad good enough to save and you can separate what you're seeing into two layers.
| Structure | Identity | |
|---|---|---|
| What it is | The hook and how fast it lands. The psychological trigger it pulls — curiosity, social proof, the quiet flex of understatement in a loud category. The composition: where the product sits, where the eye enters, how much negative space carries the line. The shape of the copy and the call to action. | The brand name, the product, the palette, the specific claims, the review counts, the founder story. |
| Why the ad worked | This layer is why the ad worked, and none of it belongs to the brand that ran it. | This layer is why the ad worked for them, and it transfers to you not at all. |
| Whether it transfers to you | Structure is craft knowledge. It has been passed between advertisers since print. | Worse than not at all: carried across, it turns into plagiarism, or into claims about your product that were only ever true of theirs. |
Almost every misuse of a swipe file is a failure to cut between the layers. The teams that copy too literally are dragging identity along with structure. The teams that copy too loosely are dropping the structure while trying to avoid the identity. The cut is the skill.
Almost every misuse of a swipe file is a failure to cut between the layers.
You can run the cut by hand
Before this is a story about software, it's a method you can use this week, with a notebook.
Take one ad from your swipe file and write a brief from it, answering only structural questions. What is the hook, in one sentence that never mentions their product? What does the ad want me to feel in the first second, and what mechanism produces that feeling? Where does my eye land first, second, third? What shape is the copy — a claim, a question, a confession, a comparison? What does the call to action assume I already believe?
Rebuild from the brief with your product, your palette, and — this is the wall that matters — only claims you can actually back.
If the reference leaned on a review count and you don't have one, the honest rebuild expresses the hook without it. The moment you invent a number to fill the slot the reference left, you've crossed from craft into fabrication, and in a regulated category you've crossed into something a lawyer gets to name.
One more rule for the hand method, because it's the one most teams skip. Be picky about which ad you brief. The screenshot you saved because it was clever is not the same as the ad that has been running for four months. An ad that is still spending to stay up is the market telling you the structure works; an ad you liked is you telling yourself. When I'm choosing what to brief, I want the boring evidence: how long has this been live? A long-running ad in a competitive category has already paid for its own test.
An ad that is still spending to stay up is the market telling you the structure works; an ad you liked is you telling yourself.
Do this ten times and you'll notice your swipe file has quietly become something better: a library of briefs. Structures you own the use of, with nobody's identity stuck to them.
We built the cut into software, so I can tell you where it breaks
At Jinn we build brand infrastructure for consumer brands, and our creative product, Vermeer, is this exact method made mechanical. I'm describing it here as a worked example, because building it taught us where the hand method fails at volume.
The extraction step is an AI that reads the image and turns a winning reference ad into a reusable creative brief — the hook, the psychology, the composition, the copy and call-to-action shape — explicitly stripped of the source brand's name, product, and claims the moment it's analyzed. What survives is the structure layer, and it is shorter than people expect. The illustrative brief we show on our own page runs to four lines: the hook is a quiet flex; the psychology is understatement against a loud category; the composition puts the product low-right and lets negative space carry the line; then rebuild it with your product, your palette, and a claim you can actually back. Read that again and notice what isn't in it. No brand name. No product. No number. A brief that short is hard to plagiarize with, which is the point.
The references themselves live in a library rather than a folder: reference ads collected, ranked, tagged, and de-duplicated, so the same winner isn't briefed five times under five screenshots. The ranking is the part I'd underline for the hand-method reader, because it's the same rule I gave you above. A reference ranks on how long it has been running. That's a plain measurement, with no model guessing which ad is good, and it's the closest thing to a verdict the market hands out for free.
The rebuild step is where honesty gets tested, so that's where we put the wall. The rebuild can only use claims verified in the brand's own record. No invented statistics, review counts, ratings, or endorsements. Our copy states the rule in one line I'll quote because it took real engineering to make true: "A missing proof point never becomes a made-up one." That rule is enforced while the ad is being made, not in a review step afterward, because a review step is a person hoping to catch a machine, and at any real volume the machine wins.
A missing proof point never becomes a made-up one.
Two mechanisms make the wall hold, and they name the two ways a hand rebuild goes wrong too. The first is where the claims come from. The rebuild reads the brand's verified record, the same record every Jinn product reads, and generation is constrained to it. If the record has a review count, the ad can use it. If it doesn't, the hook gets expressed without one. The second is what happens to the exact words. Prices, offers, and legally required lines are pulled out of the image generation entirely and stamped onto the finished frame afterward, in the brand's real font. An image model asked to paint a price will, some fraction of the time, paint something that only resembles one. Holding the exact copy out of the paint step means a price can't drift and a disclaimer can't get creative. The product gets the same treatment: it is cut out of its own photo before any image is generated, and a cut the system isn't confident in falls back to the original photo rather than shipping a mangled bottle. Notice the pattern. The parts that would embarrass you in front of a customer, the number, the legal line, the product, are handled by exact rules rather than by asking the model nicely.
Then every finished ad gets checked twice: a mechanical measurement of the contrast between the text and what's behind it, so the words are actually legible, and an AI reviewer that looks at the image and confirms the product wasn't distorted and that copy never sits over a face or the product itself. Findings, not vibes.
The reviewer's vocabulary is small and specific, and I'd encourage anyone doing this by hand to steal it. Wrong product. Product distorted. Text over a face. Text over the product. Text illegible. Scene artifact. Competitor leak. And one verdict worth earning: clean. Seven ways an ad fails and one way it passes, which is a better checklist than most creative reviews I've sat in, where the feedback was "something's off" and the fix was another round.
Two details in how the checks behave matter more than the checks themselves.
| The legibility check | The AI reviewer | |
|---|---|---|
| Measurement or opinion | Legibility is a measurement rather than an opinion. | A model grading a model is a second opinion, not a gate. |
| What it can do on its own | A hard-to-read text block can't clear on its own no matter how confident the reviewer is. | The reviewer advises; it doesn't decide. Its findings land next to the frame and a person signs off. |
Second, not every failure gets the same treatment. A cosmetic glitch in the copy earns exactly one automatic retry. A detected competitor logo earns none, because generating it again can't fix an intellectual-property problem; it stops and waits for a human. Retry the cosmetic, halt the legal: that is the whole ethic of the thing in one rule. Sets land as drafts behind a review gate, and nothing is final until someone approves it.
I'm not telling you to buy it. I'm telling you the shape of the machine, because the shape is the useful part even if you never touch our version of it.
The questions to ask any tool that offers you this
The category around AI ad creative is filling fast, and most of what's being sold is a prompt box with a template library behind it. Whatever you evaluate (ours included), the swipe-file mechanics give you the interrogation:
If the tool's output can quote the reference's claims, the cut isn't happening.
Ask specifically: if my product has no review count, what does the output do with a format that wants one? The wrong answer is a plausible-looking number.
Whether text is legible can be measured. Whether the product still looks like your product can be judged by a second AI. "Our outputs are high quality" is neither.
If the answer is yes, ask to see the loop: which result feeds back into which decision. Ours doesn't, and our page says so in plain words. There is no feedback loop from live results into generation; references rank on their own runtime and exports are recorded one way, for measurement, never to bias the next render. A tool that claims to learn from your performance and can't show you the mechanism is describing a roadmap.
The swipe file was always a library of structures. The identity was never the part worth stealing.
Vermeer is live and invite-first; self-serve signup is off while onboarding is managed by invitation.
What is the difference between being inspired by an ad and copying it?
Recreate a saved ad too literally and you've made a copy: someone else's joke, someone else's claim, someone else's brand wearing your logo. Recreate it too loosely and you lose the thing that made you save it in the first place. Almost every misuse of a swipe file is a failure to cut between the layers.
Which ad from my swipe file should I brief?
The screenshot you saved because it was clever is not the same as the ad that has been running for four months. When I'm choosing what to brief, I want the boring evidence: how long has this been live? A long-running ad in a competitive category has already paid for its own test.
How does Vermeer stop the rebuild from inventing proof?
The rebuild reads the brand's verified record, the same record every Jinn product reads, and generation is constrained to it. If the record has a review count, the ad can use it. If it doesn't, the hook gets expressed without one. That rule is enforced while the ad is being made, not in a review step afterward.
Does Vermeer learn from how my ads perform?
There is no feedback loop from live results into generation; references rank on their own runtime and exports are recorded one way, for measurement, never to bias the next render. A tool that claims to learn from your performance and can't show you the mechanism is describing a roadmap.