How to Build a Creative Swipe File You Actually Use

A swipe is a pattern you could steal the structure of, not an ad you merely liked. Before you save anything, write a one-line rule: it must have a readable first two seconds, a named hook type, and a job you could reuse on your product. Pretty ads with no stealable mechanic stay out. The file only works if every card is a prompt for production, not a mood board.

Write the rule on the first page of the doc or at the top of the Notion database so it is visible when you are tempted to dump ten TikToks after a late scroll. Allowed entries look like: question hook plus packshot, confession hook plus texture close-up, or social-proof open plus on-screen review quote. Disallowed entries look like: nice colour grade, funny comment section, or a founder film you could never recreate. If you cannot name the mechanic in eight words, you do not have a swipe — you have a screenshot.

Separate your own winners from competitor swipes in two views of the same database. Your winners already have hook rate, hold, and CPA attached; competitor ads only have a hypothesis. Mixing them without a source tag is how teams start copying vibes and stop copying the thing that actually stopped the thumb. Keep both, but never treat a stranger's ad as proven in your account until it survives your own 48–72 hour test. A competitor open that looks clever on Monday and dies at 1,000 impressions was research, not a template you owed the auction.

Fill the file from a short list: your ad account's current and retired winners, a fixed set of competitor and adjacent-category pages, and a handful of organic posts that already behave like ads. Cap a sourcing session at 25 minutes. Unbounded scrolling produces 80 near-duplicates of the same talking-head and teaches you nothing about new first frames.

Start inside your own account. Export or screenshot the ads that held a 30%+ hook rate on cold traffic, plus the ones that died when frequency hit the 2.5–3.5 band — those retired openings are still useful as 'this already burned' tags so you do not relaunch the same first line. Then pick five to eight competitor or adjacent brands and check their live ads and organic posts on a schedule, not whenever you feel behind. Adjacent categories often yield cleaner hook mechanics than a direct rival who is running the same UGC template you already use.

For each saved ad, capture the first frame as an image, the spoken opening as text, the on-screen line if there is one, and a 10-second screen recording if motion is the point. A still of second twelve is almost useless; the thumbstop lives in the first two seconds. Do not save the whole 45-second cut unless the middle contains a proof beat you specifically want to steal. Storage is cheap. Attention while reviewing is not. If you cannot reconstruct the open from the card without reopening the original video, the card is incomplete and should stay in inbox.

Build seven fields and refuse to save an incomplete card: hook type, first-frame visual, product proof, offer, platform, source, and a one-line steal. Tags are how you find three genuinely different openings when it is time to render a 6-creative batch. Folders named 'skincare' or 'good ads' collapse the moment the file has more than forty items. If a card cannot be filtered, it cannot be used on a Monday morning.

Hook type should be a closed list you actually reuse: question, confession, complaint, number or claim, pattern interrupt, social proof, direct callout. First-frame visual is equally strict: face to camera, product in hand, before mess, result close-up, text-only, motion from the side. If two cards share both the hook type and the first-frame visual, they are one pattern — keep the stronger example and archive the other. The steal line is the only free-text field, and it should read like a production note: 'open on empty bottle, then confession about wasting money on serums.'

Add a status field with four values: inbox, ready, used in a test, retired. Inbox is the dumping ground from the sourcing session; ready means tags are complete; used in a test means you already turned it into a hook in a 3 × 2 grid; retired means it lost in your account or you have burned that opening on this audience. Without status, the file grows and the useful layer hides under last month's homework. The weekly ritual in the next step exists to move cards between these four states, not to admire them.

Block 30 minutes once a week. Process the inbox, then pick three cards with different hook types and different first frames. Those three become the hook column of your next test matrix. A review that ends with 'lots of good ideas' and no named trio has failed, even if the file looks organised. The output of the meeting is three steal lines written down, not a longer archive.

The trio has to disagree with itself. One complaint, one proof-led open, and one pattern interrupt is a real test. Three confession hooks with a face in the same crop is one idea wearing three costumes, and the auction will treat it that way. If the inbox has nothing that differs on both hook type and first frame, go back to sourcing rather than forcing a batch from near-duplicates. Volume without difference is how a swipe file becomes a folder of the same ad. Write the three formula names next to the lines so a producer cannot 'improve' them into synonyms overnight.

Write the three steal lines at the top of that week's production note before you open any generator. That note is the brief: three hooks, one locked body script, two avatars, same offer and landing page. The swipe file's job ends when those three lines are written. Everything after that is production and media buying. If you skip the written trio and 'just generate some variants,' you will rebuild from vibes and the tags were theatre. Put the date on the note so next week's readout can point at the exact three cards you meant to test.

Lock one body script and one offer. Render 3 hooks × 2 avatars — six videos — from the three swipe-backed openings. Launch them together, wait 48–72 hours or about 1,000 impressions per variant, then write the winner and the kills back onto the original cards. A swipe that never gets a result tag is still a museum piece. The file only counts as used when the grid is in market.

Keep the body identical so the test reads the openings, not a pile of simultaneous changes. Name files with hook type and avatar so the readout maps to tags without detective work. In Klip Kanvas, paste the three steal lines as hook variants on one product brief and generate the two-avatar grid in one sitting rather than rebuilding each video from a blank script. The file is only 'in use' when that grid is in the ad account, not when the screenshots look tidy. If you change the offer on variant five, you are no longer testing the swipe — you are testing a new deal.

After the window, tag each source card: won (hook 30%+ and a usable CPA), grey (hook in the mid-20s, iterate the line), or killed (hook under ~20% or spend past your kill rule with nothing to show). Transfer the winning first frame and spoken line into your 'own winners' view. If both avatars lose on the same hook type, retire that pattern for this audience instead of rewriting adjectives. The swipe was a hypothesis; the 48–72 hour read is the grade. Leave the losing card visible in retired so a teammate does not 'discover' it again next month.

Once a month, archive cards you have already tested, openings your audience has seen into the 2.5–3.5 frequency band, and competitor templates that lost twice. A swipe file that never shrinks will keep feeding you last quarter's first frames. Retirement is part of the workflow, not a cleanup day you postpone until the ready layer is unreadable. Keep the retired layer searchable so you do not pay twice for the same first two seconds.

Retirement is not deletion. Keep a retired view so you can see what you already burned — especially your own winners that fatigued — and so a new teammate does not relaunch the confession hook that cooked last month. Tag the reason: lost in test, fatigued in-market, or legally or brand-unsafe to copy. Brand-unsafe copies (a rival's trademarked catchphrase, a before/after you cannot substantiate) should never have been saved; if they were, remove them from ready immediately. A retired winner is still useful as a body template; it is not useful as next week's first two seconds by default.

Use retirement to keep the ready layer small enough to scan in one sitting — roughly 30 to 60 live cards is plenty for a single product line. If ready has 200 items, you do not have a system, you have a backlog. Archive by age if you must: anything unused for six weeks goes back to inbox or retired. The file's value is the speed of picking three different openings on Monday, not the completeness of your competitive archive. If the weekly review takes more than 30 minutes, the ready layer is too fat — retire harder before you source more.