Hey there,

ChatGPT just learned to make GIFs, and I don't think that's the best part of its new update.

The feature designers should watch is saved references. Here's what's inside this week:

I gave Claude the same brief twice: a website for a run club in Paris. The first time it guessed, and the second time I prepared the project like a designer would.

Most people don't want to learn another app. They want their tools inside the chat they already use.

That's the whole idea behind the next version of Limora AI, which arrives on November 18. It's built to work inside ChatGPT and Claude, and we should get approved for an official plugin marketplace very soon.

I'm also looking at putting the full platform in ChatGPT. MagicPath took a year to reach 3 million designs, and since moving in, the founder says they've been adding 1 million a week.

We serve the same vibe coders and designers, so it's worth testing. I'll keep you updated.

We're also pushing the community side hard this time.

Stage is on the same road. Its MCP is still in testing, and I can't wait to ship it.

On Wednesday, OpenAI released ChatGPT Images 2.6.

The headline feature is GIFs. You can turn a prompt or an image into a looping animation, change its speed and edit single frames.

You can also save a reference image and reuse it in later chats, without uploading it again.

It's live on every ChatGPT plan and in Codex. Two new API models also generate images 50% faster.

Consistency has always been the weak spot of AI images. You upload the same product shot again and again, and the result still drifts.

A saved reference makes that faster. You can call the same character or product by name in every chat.

GIFs are useful too, for small jobs like loading states, animated icons and email headers.

Try it: Save a product shot as a reference in ChatGPT, then ask for a short looping GIF of it.

Read the announcement →

Free · Agent canvas
Your coding agent puts each design exploration on one canvas as a live frame, so you can compare versions side by side. Click Share and your team gets one link with comments, live cursors and voice huddles. It works with Claude Code, Codex and Cursor, and nothing goes online until you share it.

Free · Open source
Give your agent a screenshot, a mockup or a Figma frame, and it builds a real, responsive website. It then checks its work against your design until the two match, and it works with any agent.

Free · Open source
You get 76 React components for AI chat and agent interfaces, like tool calls, approvals, traces and usage. It's MIT-licensed with no Pro tier, and it installs with the shadcn CLI.

Most AI websites look generic because the images come last, and nobody plans them.

This workflow maps the page first in Claude Code. Then Limora AI generates every image in one style.

Open Claude Code, pick a design reference and paste the wireframe prompt. You get every section mapped in gray boxes, with each image slot labeled.

Upload your reference images to Limora AI and create a style. Then connect the Limora AI MCP to Claude Code.

Generate every asset in one prompt and fix any image with a quick edit. Then push them into your site, and the gray boxes become the final images.

Want to see the full tutorial? The whole workflow is in the video.

No reader build this week, so this spot is open.

Built something with AI?
Reply to this email with a link and I'll feature it here for free. Yes, for free, in front of nearly 8,000 designers and builders.

One question for you:
Which would you use first, GIFs or saved references?

Hit reply and tell me. I read every answer.

Know someone who should read this? Forward it, or send them to logiaweb.net.

See you next Monday,
Adrien