Writing
-
Prompt to Page 0.7.3: fixing the tools that fix your page
Refine with AI could change the wrong section, Fix with AI could report on a fix before it ran, and the Validation panel could show another page's errors as yours. 0.7.3 fixes them, and pages in five more design systems now look the way those systems intend.
-
Prompt to Page 0.7.2: three screen readers, and what an agent's test is worth
An AI agent tested the app with NVDA, JAWS and VoiceOver, in real Windows Contrast themes and at 400% zoom. It found real problems, and 0.7.2 fixes them. It also can't tell me the thing I most need to know.
-
0.7.0: turning the accessibility audit on ourselves
The app that checks every prototype for accessibility got audited against WCAG 2.2 AA and Section 508 itself. What was found, fixed, and is still untested — stated honestly.
-
Prompt to Page 0.6.0: a check that cannot pass should not run
MLX models could not start at all on Apple Silicon from 0.4.2 through 0.5.2 — the app looked for a filename the installer had stopped producing. A form page used to be generated twice, because the quality check flagged the back link the app itself adds. Both looked like your problem. Neither was.
-
Prompt to Page 0.5.0: a broken journey should look broken
Until this release, an exit that went nowhere was quietly rewritten onto the next page, so a broken journey still walked cleanly in the exported ZIP. The preview showed the unpatched exit all along. Only the thing you handed to someone else was smoothed over. 0.5.0 stops that, and applies the same rule to the assessment evidence document and to the quality claim the app makes about your model.
-
Prompt to Page 0.4.0: three design systems, one of them measured
The Scottish Government Design System arrives with a five-seed battery behind it. The NSW Design System and Bootstrap Italia arrive with a caveat. The app tells you which is which, and that difference is the whole release.
-
Adding a model is not the same as recommending it
Version 0.3.2 puts Qwen3.6 35B-A3B in the 32 GB picker so you can compare it with the 4-rated coder. The coder still scores higher on this job. That is the point of measuring.
-
To add a second design system, I had to find where the first one was hiding
Making the tool work beyond GOV.UK — NHS, departmental systems like HMRC, and a design system that belongs to no one — meant admitting GOV.UK was never logic in my code. It was data, hardwired in about ten places.
-
Matching an AI model to the machine in front of you
"Which model do you want?" is the wrong question to ask a user. Here's how the app answers it for them — and one UX decision I still haven't settled.
-
I wrote the counter-arguments before I wrote any code
Fine-tuning a small model to "just know" the GOV.UK Design System is a seductive idea. So I spent a day trying to kill it on paper before spending three weeks proving it in code.
-
Smaller models need smaller prompts
I assumed a weaker model needed more help. So I gave it more context — and its output got worse. Here's the experiment, the numbers, and the honest caveats.
-
Teaching a local model the GOV.UK Design System — without fine-tuning it
A model that has read the whole internet still doesn't know your design system precisely. Here's how I bring the Design System to the model at the moment it generates, behind a single sentence of prompt.
-
The small models got good — and that changes where government AI can run
gpt-oss runs in 16GB. Gemma 4 runs on a phone. The case for sending public-sector data to a data centre is getting weaker by the month.
-
I built a GOV.UK prototyping tool that never goes online
Local models are finally good enough to do real, standards-aware work on the device in front of you. So that's where this one runs.