Home / Blog / AI Workflows

Vibelets · AI How-To

How to make Meta ads with Claude and a couple of AI tools

You can run most of the path from a product page to a live Meta campaign yourself now, with Claude and two or three AI tools, in an afternoon. Here is which parts Claude does, where an image or video tool comes in, and how to launch clean on Meta.

By Siddhant, Product Marketing Manager at Vibelets

Built for people who run and create their own Meta ads

Most of the work between "I have a product page" and "my ads are running on Meta" is not the ad itself. It is the reading, the writing, the designing, and the rebuilding in Ads Manager. That loop used to need a small team. You can now run most of it yourself with Claude and two or three AI tools, and get from a URL to a live campaign in a single sitting.

This is the exact path: which parts Claude does well, where you bring in an image or video tool, and how to launch clean on Meta. At the end there is a faster way to run the whole thing in one place, and the manual version below works today and teaches you why each step matters. Name the tools as you go and you can rebuild this workflow for any store you touch.

Key takeaways

Your product page already holds the offer, the pains, the proof, and the brand look. The job is extraction with Claude, not invention.
Claude turns the page into 6 to 8 customer pains, then writes several hooks against each, so your first launch is a spread of angles rather than one guess.
An AI image or video tool renders those hooks into real creative from your product photos, so volume stops being a designer bottleneck.
Meta still wants broad targeting and enough budget to clear the learning phase. The tools change the creative, not the media math.
Claude reads the early numbers and tells you what the signals say. You keep the scale and pause calls.

Read

claude · pains

→

Write

claude · hooks

→

Render

ai · image + video

→

Launch

meta · learning

→

Read

claude · signals

Claude does the reading and writing. An AI tool renders. Meta learns. You make the two calls that matter.

01 Use Claude to read the page for angles, not features

Copy your product page text, or paste the URL content, into Claude with a simple brief: "Pull 6 to 8 customer pain points from this page. Write each as a one-sentence before-state in the customer's own words, the stuck feeling they have today."

Claude is good at this because a landing page already argues for the product. The raw material is sitting there: the frustration, the transformation, the objection, and the proof. Ask for before-states, not feature bullets. "20 grams of protein" is a fact. "The shake that actually keeps you full until lunch" is a hook, because it answers a real before-state.

This is the step people skip, and it is why so many first campaigns test one angle five ways rather than five angles once. Get the pains right and every later step has something true to build on.

02 Turn each angle into hooks with Claude

Feed the pains back to Claude and ask for hooks across different emotional levers: fear of staying stuck, the aspiration of the after, curiosity, social proof with a number, a contrarian take, a story that starts mid-scene. Ask for one punchy first line each, 15 words or fewer.

Six pains at six hooks each gives you about 36 openers before you have written anything by hand. Have Claude label each hook with its pain and its lever, so later you know why a winner won and can pull that thread into your next campaign.

Keep Claude honest here. Tell it to avoid hype and fake urgency, and to write the way a real person talks. The opener does the whole job of stopping the scroll, so the rest of the ad only has to keep the promise the hook already made.

03 Render the creative with an AI image or video tool

Now produce the actual ads. Take your product photos and the hooks into an AI image tool for statics, and an AI video tool for motion. Most of these read a reference image and a short prompt, so you feed them the product shot plus the hook and the format you want. There are several of these tools, and the workflow is the same across them: reference in, ad out.

TWO RULES

Keep the render usable

Match the first frame, or the top third of the static, to the spoken or written hook, so the promise and the visual agree. And produce both a video and an image version of the strongest angles, because the same message can win in one format and stall in the other.

Gets you: a real spread of cuts, rather than the two ads you had time to make by hand

04 Build the campaign so Meta can learn

This part stays inside Meta, and the media math has not changed. Go broad on the audience and let the algorithm find the buyers, because in 2026 the creative carries most of the targeting signal. A tight interest stack mostly starves delivery.

Watch the learning phase. Meta needs roughly 50 conversions per ad set per week to stabilize. So run 3 to 4 hooks per ad set, give each enough budget to clear that bar, and resist editing mid-learning, since a budget change over 20 percent quietly resets the phase.

You can even ask Claude to draft the campaign structure as a checklist: objective, audience, ad set count, budget per ad set, and the 3 to 4 creatives each one carries. You still build it in Ads Manager, but you build from a plan rather than a blank screen.

Rule of thumb: broad audience, 3 to 4 hooks per ad set, ~50 conversions/week to clear learning, no edits over 20% mid-phase

05 Read the first signals with Claude, then make the call

Once it is live, let it breathe, then export the early numbers and paste them into Claude. Ask it to read them in order: click-through rate and hook rate first, since they tell you which openers earned attention, then cost per result and conversion rate to confirm what the early metrics predicted.

Claude is a strong reader here. It can flag which ad set cleared the learning phase, which hook is carrying delivery, and where cost per result is drifting. What it should not do is spend your money. Keep the scale and pause decisions yours. The tool surfaces the signal. You approve the move.

The one-place version

The whole path, without the tool-juggling

Vibelets reads your store from a URL, pulls the pains, writes the hooks, generates the video and image ads, and launches the campaign to Meta. Then it flags what to scale and what to pause with the signal behind each call, and you approve the move. It is the five steps above in one place, with no card required.

Start with 50 free credits ↗ app.vibelets.ai

06 The takeaway

The path from a product page to a live Meta campaign is mostly extraction and assembly, and both compress well with AI. Use Claude to read the page for angles and write the hooks, an image or video tool to render them, Ads Manager to launch broad, and Claude again to read the first signals. Keep the two decisions that actually need you, the angles you test and the scale calls, and hand off the rest.

Frequently asked

How do I make Meta ads with Claude?

Paste your product page into Claude and ask for 6 to 8 customer pains as before-states, then several hooks against each. Render those hooks into statics and video with an AI image or video tool, build the campaign broad in Ads Manager, then paste the early numbers back into Claude to read the signals. Claude handles the reading and writing, and you keep the scale and pause calls.

Can Claude write Facebook ads from my product page?

Yes. Paste the page text into Claude and ask for 6 to 8 customer pains as before-states, then several hooks against each across different emotional levers. Claude extracts the angles that are already in the copy and writes the openers. You still render the creative in an image or video tool.

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

A language model like Claude for the pains, hooks, and campaign plan, 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 exactly what an all-in-one platform collapses into one step.

Can AI make ad creative from my product photos?

Yes. Most AI image and video tools read a reference image plus a short prompt, so you feed them the product shot and the hook and get a static or a video back. Match the first frame or the top third to the hook so the visual and the promise agree.

How many ad variants should I launch with?

Generate a pool of 30 or more from the pains-times-hooks-times-formats matrix, but run only 3 to 4 per ad set so each stays funded enough to clear Meta's learning phase.

Does detailed targeting still matter on Meta in 2026?

Less than it used to. Broad audiences plus strong, varied creative usually beat tight interest stacks, because the creative now carries most of the targeting signal.

Should I let AI scale and pause my ads?

Let AI read the numbers and recommend. Keep the scale and pause decisions human, so a tracking blip or a single good day never moves real budget on its own.

Sources

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

Vibelets · vibelets.ai