Here’s one tactical thing anyone doing AI creative can steal today.

Most of the time, the image is right and the text is wrong, or the text is right and the image is wrong. So we stopped doing them at the same time.

Let’s dive in!

The thing that pissed me off for the longest time: we'd go generate an image, the image would look great, and the font screwed up the whole thing.

Start from scratch again.

Every time. And I didn’t want to keep doing this. The image is right and the font's wrong, or the font's right and the image's wrong, so the clear solution was to do them separately.

The first thing I did was go make the images, then come back and upload that image and try to get the font right on top of it. Still wouldn't be right. And a lot of it was placement; even sliding text over a quarter of a centimeter on a 9x6 matters. It's a huge difference.

It’s amazing how much better the result is now.

Now the designer places it once, in Figma. They build the frame, drop in the real H1 copy, get the placement exactly right, and that's the only manual work anyone does. We name that text layer H1.

From there it's a custom plugin doing the work. The plugin reads the H1 layer off the file, duplicates the frame, and drops in new copy, so the new headline lands in the exact same spot, same font, same size, because it's the same layer, just with different words in it.

Claude's job is upstream of that: it writes the new headlines. The plugin takes whatever Claude generates and feeds it straight into the duplicated H1 layer. We've got it running up to 25 headlines at a time, in about five seconds.

The image was never the hard part. Turns out the challenge was making one process do two jobs at once, write the copy and place the copy.

Split those jobs up and it stops breaking.

TheLab is our own internal toolset, available for everybody. It's live now, in the Shopify App Store, $99 a month.

We built it for our own team first, we're keeping it at that price on purpose, and every tool that goes into it started as something we needed internally.

I’m not saying “make sure you use a custom Figma plugin.” There’s just a wider lesson here: when an AI workflow keeps breaking, the fix usually isn't a better prompt. AI isn’t always good at big multi step tasks, and sometime reworking your approach can really pay off.

— Torii, founder @ DREAMLABS