Home / Blog / AI Workflows

Vibelets · AI How-To

How to make 30 Meta ad variations from one product page with Claude and AI

You can build a real Meta ad test in an afternoon now. Here is how to make multiple Facebook ad variations with AI: Claude reads your product page, builds the testing matrix, and writes every variant, an AI image tool and an AI video tool render them, and you run the best few per ad set.

By Siddhant, Product Marketing Manager at Vibelets

Built for people who run and create their own Meta ads

Most people make three ads, run them, and call it a test. Three ads is not a test. It is a coin flip with extra steps, because you never gave Meta enough variety to learn what your audience responds to. The reason people stop at three is the same as it has always been: writing the fourth, fifth, and thirtieth by hand takes hours, so the volume dies before the learning starts.

What changed is that the writing and planning is now something you can hand to Claude, and the rendering is now something an AI image or video tool can do from your product photos. Claude does the reading, the matrix, and the copy. The AI tools turn that copy into real statics and motion. Meta stays where it has always been, running the learning phase on the ads you feed it. This is the whole path and which tool owns which part. There is a one-place version at the end, but the manual stack below works today and shows you why each step earns its place.

What you'll walk away with

Claude is a reader and a writer, so it does the parts that are language: pulling pains, building the matrix, and drafting every variant.
A testing matrix is a grid of pains crossed with hook types crossed with formats, which is how one page turns into 30-plus ads.
An AI image tool renders the statics and an AI video tool renders the motion, so volume stops being a designer bottleneck.
Meta needs about 50 conversions per ad set per week to exit learning, so you run 3 to 4 variants per ad set, not all 30.
Claude can also read the early numbers and tell you what they say, while the scale and pause calls stay yours.

01 What Claude does, and what it hands off

The job between a product page and a live campaign is four kinds of work: reading the page, writing the ads, rendering the creative, and launching on Meta. Only two of those are language work, and those are the two Claude is built for. It reads a page and pulls the angles, and it writes copy at volume without getting tired on the twentieth variant. Rendering and launching go to tools built for them. So the split is clean: Claude reads and writes, an AI image tool and an AI video tool render, Meta runs the auction and the learning phase. Naming who owns what up front keeps a DIY stack from turning into a mess, because you stop asking one tool to do a job it was never good at.

Claude

read · matrix · copy

→

AI image + video

render statics + motion

→

02 Have Claude build the testing matrix

A matrix is multiplication done on purpose. You pick three axes, pains, hook types, and formats, then cross them, and the ad count fills itself. Six pains at three hook types at two formats is 36 distinct ads before you have written a single one by hand. That grid is the difference between guessing and measuring, because every ad maps to one angle and one hook, so a winner tells you why it won.

Claude is the right tool to build the grid because every axis is language. Paste your product page text or its URL content into Claude and ask it to pull 6 to 8 customer pains as one-sentence before-states in the customer's own words. A page already argues for the product, so the frustration, the transformation, and the proof are sitting in the copy waiting to be pulled. Ask for before-states, not feature bullets. "20 grams of protein" is a fact. "The shake that keeps you full until lunch" is an angle.

6 pains

before-states

3 hooks

fear · aspiration · proof

2 formats

image + video

6 × 3 × 2 = 36 variants. Claude fills the grid, you pick what runs.

03 Write every variant with Claude

Pull the angles from the page

Give Claude the page and a plain brief: "Pull 6 to 8 customer pains from this page. Write each as a one-sentence before-state in the customer's words." Ask it to label each pain so you can trace it later. This is the axis most people skip, and it is why so many tests are really one angle worn five ways.

Claude gives you: the rows of the matrix, in the customer's words

Cross each pain with hook types

Feed the pains back to Claude and ask for hooks across different levers: fear of staying stuck, the aspiration of the after, curiosity, social proof with a number, a contrarian take. One first line each, 15 words or fewer, no hype and no fake urgency. Six pains at three hooks gives you 18 openers, each tagged with its pain and its lever.

Claude gives you: the columns, so each angle gets a fair shot

Draft the full ad for each cell

Now ask Claude to write the body under each hook: primary text, headline, and the promise the hook made. Keep it to the same voice across all of them so the test measures the angle, not the tone. This is where the matrix goes from a list of openers to a bank of finished copy you can hand straight to a rendering tool.

Claude gives you: finished copy for every cell in the grid

Write the render briefs too

Claude can also write the prompts for the image and video tools: the scene, the first frame, and the on-screen text that has to match the hook. Ask it for a short prompt per variant that names the format and the product shot to use. Now you walk into the rendering step with a brief for every ad instead of a blank prompt box.

Rule of thumb: 6 to 8 pains, 3 hook types, 2 formats, and Claude drafts copy plus a render brief for each

Claude gives you: render-ready prompts, one per variant

04 Render the statics and motion with AI tools

With the copy and the briefs done, the rendering is mechanical. Take your product photos and Claude's briefs into an AI image tool for the statics and an AI video tool for the motion. Most of these read a reference image and a short prompt, which is exactly what you now have for every variant. This is the part that used to cap people at three ads, and it is the part a mix of AI tools removes.

Two rules keep the output usable. Match the first frame of the video, or the top third of the static, to the hook, so the promise and the visual agree. And render the strongest angles in both image and video, because the same message can win in one format and stall in the other. You do not need all 36 rendered at once, just a pool with real range for the matrix to draw from as cuts win and lose.

05 Run 3 to 4 per ad set for the learning phase

Here is where volume meets the media math. You generated 30, but you do not load all 30 into one ad set. Meta needs roughly 50 conversions per ad set per week to exit the learning phase, and more ads in one ad set splits the budget so thin that no single ad ever reaches that bar. So you run 3 to 4 variants per ad set, drawn from the larger pool, and let each one get enough budget to prove itself.

The pool

30-plus variants from the matrix
Every pain and hook covered
Held in reserve, tagged by angle
Fresh cuts ready to swap in

The ad set

Run 3 to 4 variants at a time
Broad audience, creative leads targeting
Budget clears ~50 conversions/week
Rotate winners up, cut losers, refill

The pool is your bench of tested angles. The ad set is where 3 to 4 of them compete on enough budget to reach a real read. As cuts win or lose, you refill the ad set from the matrix, which is how a test stays alive past week one instead of ending when the first three ads flop.

Once it is live for a few days, close the loop with Claude again. Export the early numbers and paste them in, and ask it to read click-through and hook rate first, then cost per result to confirm. Claude is a strong reader of that table. What it should not do is spend your money, so the scale and pause calls stay with you.

The one-place version

Skip the tool-juggling

The manual stack works, and it also means Claude, an image tool, a video tool, and Ads Manager in four tabs. Vibelets does the whole path in one place: it reads your store from a URL, builds the matrix, generates the 30-plus video and image variants, and launches the campaign to Meta. It flags what to scale and pause with the signal behind each call, and you approve the moves.

Start with 50 free credits ↗ app.vibelets.ai

06 The takeaway

Making 30 Meta ad variations from one page comes down to who does which part. Claude reads the page, builds the matrix of pains and hooks and formats, and writes every variant plus its render brief. An AI image tool and an AI video tool turn those briefs into real statics and motion. Meta runs the learning phase on 3 to 4 variants per ad set, drawn from the pool. Keep the two decisions that need you, the angles you test and the winners you scale, and hand off the rest. That is a real test from one product page, in an afternoon.

Frequently asked

How do I make multiple Facebook ad variations with AI?

Have Claude read your product page, pull 6 to 8 pains as before-states, cross them with a few hook types, and write every variant. Then render the statics with an AI image tool and the motion with an AI video tool. That turns one page into 30-plus variations without designing each one by hand.

Can Claude write Facebook ad copy?

Yes. Paste your product page text or URL content into Claude and ask for pains as before-states, hooks across different levers, and the full ad body under each hook. Claude extracts the angles already in the page and drafts the copy at volume, which is the slow part when you do it by hand.

What AI tools do I need to make Meta ads myself?

Claude for the pains, the matrix, the copy, and the render briefs, an AI image tool for statics, and an AI video tool for motion. Ads Manager is where you launch. That is three or four tools stitched together, which is what an all-in-one platform collapses into one step.

Can Claude build a creative testing matrix?

Yes. Ask Claude to cross three axes: 6 to 8 pains, a few hook types each, and both formats. It returns a grid where every variant maps to one angle and one hook, so when an ad wins you know why it won and can carry that into the next campaign.

Does Claude make the ad images and videos?

No. Claude handles the language: the angles, the copy, and the prompts for your creative. The actual images and videos come from an AI image tool and an AI video tool that read your product photos and Claude's briefs. Keeping that split clean is what makes the stack work.

How many ad variations should I run per ad set?

Run 3 to 4 variants per ad set. More than that splits the budget too thin for any single ad to reach the roughly 50 conversions per week Meta needs to exit the learning phase, so the ad set thrashes instead of learning.

How many ad variations should I make from one product page?

Generate a pool of 30-plus from the pains-times-hooks-times-formats matrix, then run 3 to 4 per ad set from that pool. The large pool gives you range across angles, and the small run keeps each ad set funded enough to learn.

Sources

Meta learning phase and ~50 conversions per ad set: facebook.com/business/help
Broad targeting and creative-led delivery in 2026: practitioner benchmarks
Creative volume and testing-matrix structure: 2026 industry reporting

Vibelets · vibelets.ai