• 100 School
  • Posts
  • 💯 How to (finally) design with AI

💯 How to (finally) design with AI

Plus: what GPT-5.5 got better at, where Opus 4.7 still wins, and the part of your job that's about to get a lot more valuable.

Author

Max Haining
April 26, 2026

I gave ChatGPT our logo last week and it gave me back an entire brand kit.

Hoodies that say BUILD THE FUTURE TOGETHER in our red. Water bottles. A Lego set. A tiny pixel-art handheld console with "LEVEL UP" on the screen.

None of which we make. All of which look like we could make them tomorrow and nobody on the team would blink.

Prompt: “Can you create a visual brand and swag kit for 100 School. Here’s our website, logo, and brand color palette.” + attached reference image.

The model is ChatGPT Image-2, which OpenAI shipped with a leaderboard score that took the largest lead in Image Arena history. The reference was from Justine Moore who built a brand identity for a16z from a logo and a URL.

These past two weeks, three frontier models landed in single stretch: Image-2, Claude Design on Opus 4.7, and GPT-5.5. The thing I've been thinking about all weekend isn't any one of them on its own but what they all add up to on a Monday morning when you're trying to do your actual job.

Window into the Future 🔮 

Back to the brand kit for a second because the part that surprised me wasn't the polish.

It was that Image-2 didn't make us a logo. It read what we already are (the mountain, the red, the copy from our website..) and produced things that don't exist yet but definitely could. That's a different operation than image generation as we've known it.

Claude Design is doing the same shift for structure that Image-2 is doing for visuals.

Peter Yang, a product manager, used Claude Design to build a video, slides, a website, a mobile app, and a design system in 16 minutes. None of them were perfect but they all looked shippable.

Shreya Heda, a designer, redesigned a brand she admires from scratch in 15 minutes and paused mid-video to ask: "will portfolios even matter after this?" Then answered her own question that maybe the next portfolio is just handing someone a brief and saying "give me a week, I'll show you what I can build."

If you want to actually be good at Claude Design, Ryan Mather, an Anthropic designer, posted seven tips this week. The three that matter if you're not a designer are

  • don't describe edits with words, point and click using the Comment tool (saves 40% of your token budget)

  • set up your design system in the first hour, so everything generated after is on-brand without re-specifying

  • and prompt from your actual work context.

Then GPT-5.5, which ties this together. Tom's Guide ran Opus 4.7 against it on seven hard reasoning prompts. Claude swept all seven. But the more honest comparison came from Every's Vibe Check, where they called GPT-5.5 "the model I'd use when I need to get the job done without babysitting it." That's a comparison of friction. And friction is the thing that's been disappearing across all three launches.

A take that Tommaso Nervegna put well in his Substack, which now applies to all three at once: when everyone uses the same AI, the output starts to look the same. What stands out are the decisions AI can't make for you like cutting a section even if it looks "right," breaking a pattern because your users need something different, and leaving space when the AI would keep adding.

When three frontier models all get good at the same things in the same week, what stays expensive is the part nobody can outsource: which model for which task, which sentence to keep, which client problem is worth solving. That's the job now. And it's the work we keep building 100 School around because every team I talk to has the tools, the budget, and the buy-in. What they don't have is somebody to sit with them and say here's where this goes in your actual week.

Your experiment for this week 👇

If you only do one thing this week, do this: take 30 minutes, pick one of the three (Image-2, Claude Design, GPT-5.5), and try it on a real thing you have to make.

  • Image-2 👉 feed it your brand assets, ask for a kit. Or try the menu test from TechCrunch.

  • Claude Design 👉 one sentence first. Let it ask the eight questions. Use Ryan's Comment tool tip for edits.

  • GPT-5.5 👉 give it a messy input you've been avoiding to turn into a clean output (a transcript, meeting notes, a pile of survey replies)

Reply with what you made. I'll feature the best ones in next week's issue.

The 100 School brand kit at the top of this email took me one prompt and about ninety seconds. The impressive part is that it's Sunday night and I'm still thinking about what to do with it.

Join our free Claude workshops next week 🎟️

You’ve probably saved a bunch of Claude demos by now. You tried a few things then maybe dropped off. It happens.

If you want to actually get something working, this is the easiest way to start.

Harold and Kieran are running two live workshops this upcoming week. You’ll build something you can use (not just watch along) in just 4 hours.

You don’t need to know what you’re doing yet. You’ll figure it out as you go.

Before you go ✌️

What's become your go-to for design work with AI? Do you still use Figma?

Also, don’t forget to reply with what you’ve built. I’d love to see it!

See you next Sunday!

Max 

P.S. Want to make your team & company AI-first? Let us help here.