Notes on design, on building with AI, and on the things that only show up once you have shipped them.
7 August 2026 · 6 min read
I stopped describing my product to AI. I just gave it the HTML.
HTML runs both ways. The most valuable HTML isn't what the AI writes, it's your real product fed back in. That's what makes AI design honest.
7 August 2026 · 4 min read
I gave the AI my design process, not my design decisions
Automation was never the goal. The best AI in design does everything around the judgment and none of the judgment.
I turned my design process into a dozen small skills, not one giant prompt
A workflow is skills plus handoffs plus checkpoints, not one prompt. Cut it into roles with a human gate between each.
I built a free tool that gives your AI your real product
Your AI gives generic design help because it has never seen your product. The fix is a free tool that gives it your real product.
7 August 2026 · 3 min read
I taught the AI to push designs into Figma. Then it quietly stopped.
An AI that silently does the lesser thing is more dangerous than one that errors. Verify the capability, don't trust the connection.
We rebuilt our Figma design system out of the code, not the other way round
Generate the design artifact from the source of truth, then have a human reconcile. The code is the source of truth.
We ran our design process on itself, and it broke in useful ways
The only honest test of a workflow is to use it for real. Dogfooding surfaces the failures no design review will.
How do you teach an AI your taste?
Capture preferences as promotable rules, not one-off corrections. And guard against over-fitting to a single offhand comment.
It worked great on the smart model. Then I ran it on a cheaper one.
Don't rely on the model being smart. Make the instructions load-bearing, so a weaker model can't wander off.
The AI told me our brand colour with total confidence, and it was wrong
The AI writes the extractor, not the extraction. Ground truth comes from a script, never from the model's memory.
A screen list lies. I found a whole feature I didn't know we shipped.
A screen list hides states, personas, and transitions. Model it as a graph, source it from code, and remember that absence of observation is not absence in production.
Our documentation updates itself, because updating it is the AI's job
Documentation dies because updating it is nobody's job. Make maintenance a step in the workflow, not a chore after it.
I sat down to turn my workflow into a swarm of agents. Most of it refused.
Parallelise the gathering, keep the judgment linear and gated. Tier models by judgment vs fetching, not seniority.
Every modal in our app shared one URL. Our analytics couldn't tell them apart.
If you can't address it with a URL, you can't measure it. Give every modal its own URL, and watch out for the gotchas.
My AI's rules only worked because one tool bothered to read them
An instruction is only as good as whether the host loads it. Portability is reachability, not naming.
Six designers, one GitHub account, and a script that saves us from ourselves
Don't enforce process with reminders. Enforce it with automation that fails loudly.
"Out of scope" usually means "we'll repaint this in six months"
Say no in two flavours. Design the near-future in, even when you don't build it.