How to Automate AI Ad Creation With Claude (A Self-Improving System)

You might be making AI ads one prompt at a time. Write a prompt, generate, look at it, adjust a word, generate again, and after forty tries you have one clip you can use. That is how I worked for two years, and it does not scale past one project at a time.

This is the setup I use now. Claude runs the generations, keeps the records and reads my feedback before it starts the next batch. I walked through the whole thing on camera in May 2026. This post covers what the system is, why it works, and what to expect when you build your own.

The system has four parts, and none of them is a clever prompt

• References. Your people, your products and your locations, prepared once so the model has something to stay faithful to.

• Memory. Claude records what it used and what it wrote for every single generation, so nothing is guessed twice.

• Rules. A short set of instructions for your project, written in plain sentences, that Claude reads before every batch.

• Feedback. A sheet where you mark each output approved or rejected and write a line about why.

The prompt writing still happens. Claude does it, from a framework I have refined over two years of client work, using the four parts above as context. That is the difference between an output that looks like your campaign and one that looks like stock.

Claude only needs to see your references once

Most people upload the same product photo into the prompt bar every time they generate, and the results drift anyway. The fix is that Higgsfield stores every uploaded reference under a unique identifier, and Claude can use that identifier in a prompt in place of the image.

So the references get uploaded once at setup. After that every prompt Claude writes points at the identifier for the right character, the right product and the right location. Same model wearing the same trainers in the same world across a batch of thirty, without anyone re uploading anything. Identifiers are tied to your own account, so nobody else can use yours.

The feedback loop is what makes it improve

After a batch I open the sheet and go through the rows. Approved, rejected, or a note. The notes are specific: the bottle is too large next to the model, the end screen is wrong, the camera move is too fast for the line.

Before the next batch, Claude reads the sheet. It leans toward the shots I approved and away from what I rejected, and it stops making the mistakes I wrote down. After three or four batches the outputs look like my taste rather than the model's default. Nobody has to retrain anything. It is just a spreadsheet Claude reads.

A batch is one instruction

Here is a real one. Ten fifteen second videos with Seedance, model one and model two in location sixteen, varied camera moves, sixteen by nine, use the perfume product, they find the bottle somewhere in the scene, end on a branded end screen, sound effects only and no music.

Claude checks the references and the rules, confirms the model, tells me what the batch will cost in credits, and starts generating. Failed jobs get resubmitted and logged. Every video lands in a folder with its full prompt recorded in the sheet before the first one has finished rendering.

Stills work the same way. Thirty images in location fifteen, half noir and editorial, half with motion blur and depth, model two only, nine by sixteen for stories. It ran while I made a coffee.

The first batch will be about seventy percent right

Expect that. In mine the end screens were the weak point and the bottle was oversized in two shots. Neither needed a rewrite. Both went into the notes column and the next batch fixed them. Treat the first run as calibration. The point of the system is that the second run is better than the first without you doing more work.

You can also run it overnight, or on a schedule, so a batch is generated and filed before you sit down.

What it costs

A Claude subscription and Higgsfield credits. Claude reports the cost of a batch before it runs, so there are no surprises. Seedance 2.5 generations are expensive, so test at low resolution and render only the keepers at full quality. The walkthrough was recorded on Seedance 2.0. We run the same setup on 2.5 now and nothing changes.

What I keep back

The video shows the system running. It does not hand over the folder that runs it: the prompt framework, the reference and rules templates, the logs and the handoff document that tells Claude how to use them. Those took two years to get right and they are available as a product if you want to skip the building. You can build your own version from what is above. It will take you a few weeks to get it to the same place.

Common questions

Do I need to know how to code?

No. Setup is a few pastes into Claude Code, and when something fails Claude tells you what the options are. I learned what a command line tool was two weeks before recording the walkthrough.

Can it use my own products and people?

Yes. Your references go in once and are used by identifier from then on. That is the whole reason it holds consistent.

Which models does it use?

Nano Banana Pro for images and Seedance for video, both through Higgsfield, as of September 2026. Any model available on the platform can be called.

How does it get better over time?

Through the feedback sheet. Your approvals, rejections and notes are read before every new batch.

Does it only make ads?

It makes whatever your references and rules describe: editorial stills, product shots, story sequences, social cutdowns. Ads are where the volume pays off.

If you want to see it running and ask the people who use it every week, join the community.