11-16-26
11-16-26
AI is faster at doing design than me, and that’s simply a fact. It still gets plenty wrong, but I’m using it every day, quickly and proficiently, and I’m building a tool that matches the speed of my workflow to help me go even faster.
Developers aren’t handwriting code anymore, and designers won’t be hand drawing every rectangle either.
Last time I wrote about building the /skill page entirely in HTML. That page took 36 commits, and since then I’ve been building something bigger with the same method, and the last eight days alone have 637 commits in them.
Fig. 1
HTML is the canvas
HTML is the building block of the internet, and canvas design has always been drawing pictures of what the HTML should ultimately look like.
I still adjust things by hand all the time, and there are still plenty of times you need to dial something in with a precise touch.
But generating tons of HTML is super fast. I generate, organize and decide in HTML, and go to the canvas when I need to, so building in straight up HTML feels like the fastest way to design right now.
You don’t have to learn how to code to work this way. It’s still true even though we’re working in real code, because Claude, ChatGPT or Codex is drawing the rectangles. Even when you draw rectangles in Figma, there’s code running in the background that was written ahead of time to let you do that, and now AI can generate that code. I still use the canvas when I need it. This is just a very fast way to work.
Keep 5 percent
I’ve used ChatGPT since it came out, Cursor as soon as it came out and Claude Code as soon as it came out, and I’ve been trying to stay at the tip of the AI spear. For the longest time, I didn’t think it would get good at design, and I sort of didn’t want it to. Now it can do a lot right.
There’s something unsettling about AI creeping in and getting good at what I’ve been doing professionally for 20 years. It’s unsettling for almost everyone to have their career disrupted by new technology, and it’s happened over and over. Radio did it, then television, movies, DVDs, the iPod and Spotify, and long before all of them, the sewing machine. It just hadn’t happened to us yet, and now it is. What’s helping me get past it is learning the new tool, being adaptive and changing.
So I explore tons of concepts, throw 90 or 95 percent of them away, and keep the 5 percent worth keeping. I keep building context about my design decisions as I go, and keeping that momentum going is really helping my design work.
How do I pick the 5 percent? I look at whether it’s good or not. When something about one of them is interesting, that becomes a new thread for me to pull. It takes skill to know which one to keep, and it also takes a point of view and a little bit of conviction. You could easily keep 20 percent because some of it looks okay, but usually only 5 percent makes it past my filter as something that’s great, or could be great with more fine tuning.
Fig. 2
The decision stays mine
I never hand off the actual decision. Claude can generate as much as it wants, but I put in the inputs, I decide the outputs, and I refine, and I refine.
I’m not interested in a one shot. If something can be one shot, then the creative part is the idea and the input you gave it, and that’s creativity too.
And plenty of times you go down the wrong trail and you’re just spinning your wheels. When that happens, throw a stick in the spokes and switch gears completely, grab some different inspiration, or ask for something completely different.
Talk. Point. Paste.
For the last year, maybe longer, I’ve been taking screenshots and pasting them into my AI agent. I’ve been recording long voice notes, and I’ve been working on JSON files mapped to words so I could do faster video edits. It all felt fast, but also sort of slow at the same time.
Fig. 3
All of it was about building context for my agent, with the words, the screenshots, the recordings and the tools I’ve been using and building. I keep looking for more and more ways to be efficient with AI.
After a conversation Ridd and I had about our workflows and how they were evolving, I got inspired to build my own tool.
It’s called FloatPrompt, and it’s a Mac app in beta.
You record your voice and point at things on your screen, and you can draw on it too. It saves everything in a very slim Markdown file, along with your session data and screenshots. You paste the path to your agent, and your agent knows exactly what you were talking about.
Fig. 4
I’m using it exclusively to work with AI now, drawing on my screen, pointing at things, recording long transcriptions and pasting them straight into my agent. It’s completely replaced the dictation apps I was using before, and one day this week I used it 310 times.
Restraint
The onboarding went through 17 rounds of concepts because I was trying to shove too much into it. I’d lost track of the marketing site and everything that happens before someone downloads the app, so I cut a lot of it and simplified.
Fig. 5
Fig. 6
Restraint is one of the Shift Nudge design principles I try hardest to follow. What’s the minimal amount of information I can share to still get the job done? Go too hardcore, and it just becomes confusing.
Absorb the friction
The more friction you can absorb in the design process, the less friction you hand to the person using it. Really good product design makes things slippery, easy to use and easy to set up, and that’s why I spent so long on the onboarding, the installer window, the animations and the sounds.
Fig. 7
The tiny details
Design is all about sweating the details, so literally design every detail. If you don’t, you’re leaving it up to chance, and by definition it’s not designed.
Also, if I’m going to be looking at something every single day, hundreds of times a day, I want it to be as perfect as it can be.
And there are certain things that just need to happen in any app. If you’re pressing a button to talk, you need to be able to talk immediately, so it needs to feel instant. If I have to wait a few seconds, it’s not going to be worth it.
So everything we do has to make it fast, because speed will never not be a wanted attribute. Right now, letting go of the key puts your words on the page in about a seventh of a second.
The little ghost on the capsule is about 15 points tall. When you have white shapes with small black objects, they can get lost, and when things get really small, they can look different, so you have to adjust optically. I judged the ghost’s eyes at 28 pixels, magnified pixel for pixel, and they ended up 10 percent bigger than the logo’s eyes, in every pose.
Fig. 8
You do things by math first, and then you refine it optically. That’s the same across all design.
The capsule is full of those adjustments. A bare × icon or check icon leaves more room around it than your eye expects, so whatever follows it starts a few points early.
When you’re fine tuning with sliders, you can see changes instantly. You just move the slider back and forth while you’re looking at the object, and you might not even know exactly where you are on the slider, but once you see it, it’s like, oh yeah, there it is.
Fig. 9
Designing for agents
Designers will absolutely design for agents as users. Many already are, and if you’re not, I’d encourage you to start thinking about it.
That Markdown file FloatPrompt saves is written for an agent to read, and I’ve gone through hundreds of rounds of what to keep in it. I’ve been cutting and trimming and leaning out the capture data as much as possible, while keeping as much detail as possible, so the context is exquisitely precise.
Fig. 10
Designing for an agent looks a lot like designing for a person. Be super clear, give great context and be a great copywriter. Be methodical and systematic, and think through every scenario that could go wrong, from the simple version to the medium version to the really complicated one.
That’s good information architecture and good user experience design, and it matters for users and for agents.
Heading into 2027
If I were trying to improve my design and AI skills heading into 2027, I’d learn everything I could about good design and learn to wrangle AI agents into helping me build. I’d get super comfortable generating HTML and working in actual code with agents, while making really good design.
That just so happens to be what I teach in Shift Nudge, in the interface design training and in the Claude Code training for building with AI. It’s the direction everything is headed for the new material I create. Apply here if you want a head start on 2027.
Try this
One thing to try this week is to generate 20 completely different ideas in HTML, using your own adjectives and your own references. Pick one, then ask for 10 versions with more subtle changes, and see how far you can go, just as a throwaway experiment.
When I did this for the ghost, the first rounds of ghost marks only found the direction. After that came lots and lots of revisions and refinements. The shape, the tilt, the eyeballs, the little skirt wave, the crops inside the circle, the eye animations, and custom shader adjustments on the app icon. I went through more than 100 app icon variations in Icon Composer alone. That took a lot more than one shot.
Fig. 11
Fig. 12
What’s next
I’ll be sharing the FloatPrompt beta with active Shift Nudge members as soon as possible.
Happy designing,
Practical design wisdom delivered weekly.