AI Ad Creatives
4 min read
Testing 5 random ads teaches you nothing. When the winner differs from the losers by hook and angle and format all at once, you know something won but you have no idea what, which means you can't make another one.
The free Claude skill hook-testing-matrix builds a grid where every cell answers exactly 1 question: hooks on one axis, formats on the other, angle held constant.
This post covers how the matrix is sized, why creative tests are judged on attention metrics rather than CPA, and how the winners get routed afterwards.
Key takeaways
Default shape is 3 hooks by 2 formats with the angle held constant, which gives 6 cells that each isolate 1 variable.
Each cell needs at least 2,000 impressions before it is judged, so the budget is sized to the impression floor rather than to a hunch.
Tests run 3 to 5 days with every cell launched simultaneously, so all cells face the same auction conditions.
Judging happens on attention metrics: hook rate, hold rate, and CTR. A 3 day test cannot judge CPA and this skill never pretends it can.
1 variable per axis
The discipline is the whole product here. Hooks vary down the rows, formats vary across the columns, and the angle stays fixed for the entire matrix. When a cell wins, the reason is structurally obvious, because only 1 thing about it differed.
Change the angle too and you've built a test that produces a winner and no lesson. You'll get a number, you'll scale the ad, and 3 weeks later when it fatigues you'll be back at zero with nothing transferable.
Attention metrics, not conversions
Metric | Pass mark |
|---|---|
Hook rate, 3 second views over impressions | 30% or above |
Hold rate, ThruPlays over 3 second views | 25% or above, video |
Link CTR | 1.0% or above, with relative ranking mattering more |
This is the part people argue with, so the skill is explicit about the reconciliation. Conversion verdicts still belong to the scale hold kill skill, after the learning phase has completed. A 3 day test at kill fast budgets cannot produce a trustworthy CPA, and pretending otherwise is how good creative gets buried by small numbers.
The matrix answers which creatives earn a conversion budget slot. The conversion campaign answers whether they deserve to keep it. 2 questions, 2 metric sets, no contradiction.
Fill the empty cells before the test runs
A matrix only works if every cell has an asset in it. Vibelets generates on-brand product ads from your store in minutes, so the gaps get filled without a production bottleneck.
Sizing the cells honestly
Cell budget works out at roughly 2 times your account CPM, which is what it takes to clear 2,000 impressions. The skill computes this with your actual CPM and totals the test cost.
When the total exceeds what you can spend, the rule is to cut cells, not impressions per cell. A 4 cell test read properly beats a 6 cell test where every cell is judged on 900 impressions and half the ranking is noise.
Everything launches at once in a single test campaign, kept separate from scaling campaigns, with creatives named by product, angle, hook number, format, and version so the read is unambiguous.
How to set it up in Claude
Download the skill file from the skill's Google Drive folder, free, no email gate, then add it to Claude by uploading it as a skill under Settings or attaching it to a project or chat. Anthropic's Help Center guide to skills covers the upload path.
Bring your assets or briefs, plus the daily budget you can afford to spend on learning, which is separate from your scaling budget. The skill lays the matrix, flags empty cells that need production, computes the cost, and sets the read date.
On the read date, cells get ranked, hook winners get declared, and the winning pattern becomes the brief for the next batch rather than a one off result.
Conclusion
A test that changes everything proves nothing. Hold the angle, vary 1 thing per axis, fund each cell to its impression floor, and judge on attention. Then you'll know why the winner won.
Build a test that actually teaches you something
Free Claude skill, yours in 1 click.
Frequently asked questions
How many creatives should I test at once?
The default matrix is 3 hooks by 2 formats with the angle held constant, giving 6 cells. What matters more than the count is that each cell differs by 1 variable and clears its impression floor.
How long should a creative test run?
3 to 5 days, with every cell launched at the same time so they face the same auction conditions. Never judge a cell before it clears its impression floor, and never let a day 1 peek kill a cell.
Why judge creative tests on hook rate instead of CPA?
Because a 3 to 5 day test at kill fast budgets cannot produce a reliable CPA. Attention metrics are leading indicators available at that volume. Conversion verdicts come later, after the learning phase, from the scale hold kill skill.
What is a good hook rate on Meta?
The skill uses 30% or above as the pass mark for hook rate, 25% for hold rate on video, and 1.0% for link CTR. Relative ranking between your own cells matters more than hitting an absolute bar.
What happens to the winning creatives after the test?
Winners route into a conversion campaign through the campaign architect skill, and the winning pattern becomes the brief for the next production batch, so each test compounds into the next.
Sources
Make ads that sell
your Products.
Generate on-brand product ads in minutes, no designer or brief required.
Trusted by 400+ growing brands