It's Monday. Here's what changed in AI this week, and what it means for your design work:

I'm writing this from an airport, waiting on a flight to Cyprus. It's the last morning of a two month break. I don't really want it to end and I'm desperate to start at the same time.

Most of that time was France, with a lot of moving around Europe in between. I sat down properly with family I hadn't seen in far too long. I caught up with old friends who I usually only ever text. I also met a few people completely by accident, and they gave me some of the best conversations I've had all year. Two months ago I needed that more than I was willing to admit.

Then something flipped last week. I stopped enjoying the rest and started getting restless inside it, and that's usually the signal.

Q4 starts in a few weeks and we have solid objectives to reach before the end of the year. Now is the best possible time to lock in. I'm going back to a much bigger rhythm.

Two years of chasing photorealism, and the update that actually changes your workflow turns out to be a file format.

OpenAI added a background="transparent" parameter to GPT-Image-2. Set your output to PNG or WebP and the model builds the alpha channel while it generates, instead of you cutting the background out afterwards.

That difference matters more than it sounds. Background removal has always broken on the same things: glass, smoke, hair, fine strands, soft shadows. Because the model now knows it's making a cutout from the first pixel, those edges come out clean.

And it's live in Limora too. Every icon, illustration, decorative shape and logo variation you generate can come out fully transparent, ready to drop straight into Figma or your codebase. Two reasons to do it there instead: with GPT-Image-2 you have to fight the prompt, because if you write "a bottle on a marble table" and ask for transparency you'll still get the table, so you have to spell out "isolated object, no backdrop, no rectangle" every single time. Limora's icon and illustration workflows handle that for you. And Limora applies your Brand Kit on every generation, so a set of twelve icons comes back in your colors and one consistent style instead of twelve unrelated objects you then have to reconcile by hand.

Start with the thing you'd normally redraw or outsource: a set of icons for a landing page, or the little spot illustrations for an empty state. Generate the full set in one pass so the style holds, then check the edges at 400% before you ship, because soft shadows are still where these models cheat. If you want the Brand Kit applied automatically, do it in Limora.

Every time you need to explain how a system fits together, you end up in the same place: a generic flowchart that looks like it was made in 2011, or an hour lost in a diagramming tool that wants your credit card first.

FossFLOW is a free, open source tool for building those beautiful 3D isometric diagrams you've seen in infrastructure docs and architecture decks. It runs entirely in your browser as a PWA, works offline, saves as you go, and ships with a big extensible icon library covering AWS, Azure, GCP and Kubernetes. Nothing leaves your machine, and the whole thing is MIT licensed on GitHub.

Most AI-generated screens look generic for one reason, which is that the model invents its own components instead of using yours. Claude Design has a sync command that fixes exactly that, and most designers haven't touched it.

💌 Know a designer who should be using AI smarter?
Forward them this email. Or just send them to logiaweb.net to join.

See you next Monday,

Say hi on X, Instagram or YouTube