Notes from building Prompt to Page.
Working notes on local models, the GOV.UK Design System, accessibility, and what it takes to keep a prototyping tool local-first.
- 7 min read
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.
Read post - 7 min read
Prompt to Page 0.7.2: three screen readers, and what an agent's test is worth
An AI agent ran NVDA, JAWS and VoiceOver against Prompt to Page. 0.7.2 fixes what it found, from questions JAWS couldn't hear to a preview 4 pixels tall at 400% zoom. Here's what that test is worth, what it isn't, and how the release held itself back once.
Read post - 9 min read
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.
Read post - 20 min read
Prompt to Page 0.6.0: a check that cannot pass should not run
Version 0.6.0 fixes the MLX startup failure on Apple Silicon, removes a quality rule that failed the back link the app itself requires, and stops a corrective pass from shrinking a page into a fragment.
Read post - 18 min read
Prompt to Page 0.5.0: a broken journey should look broken
0.5.0 stops the export smoothing over a broken journey, applies the same rule to the assessment evidence document and the per-model quality claim — and ships the eighth design system stamped Not measured.
Read post - 10 min read
Prompt to Page 0.4.0: three design systems, one of them measured
Version 0.4.0 adds the NSW Design System, Bootstrap Italia and the Scottish Government Design System. Only one has a five-seed, per-model quality battery so far — and the picker says so.
Read post - 9 min read
Adding a model is not the same as recommending it
Qwen3.6 35B-A3B is in the picker from 0.3.2 — rated a 3, sitting next to a 4-rated coder. Why shipping a model and recommending it are different promises.
Read post - 8 min read
To add a second design system, I had to find where the first one was hiding
Adding NHS, HMRC and a genuinely neutral design system forced me to notice that GOV.UK was never logic — it was reference data, quietly hardwired across the codebase.
Read post - 5 min read
Matching an AI model to the machine in front of you
How the app picks a model that fits the user's hardware — and why I'm still unsure whether to hide incompatible premium models or show them, greyed out, with a reason.
Read post - 6 min read
I wrote the counter-arguments before I wrote any code
Why I wrote the counter-case first — and the go/no-go gates that keep a seductive ML idea from becoming an expensive mistake.
Read post - 6 min read
Smaller models need smaller prompts
Trimming worked examples from the prompt took a small model from 1/7 passing test cases to 4/7 — and the average score from 59 to 83.
Read post - 5 min read
Teaching a local model the GOV.UK Design System — without fine-tuning it
How Prompt to Page uses a class list and just-in-time component retrieval to keep a local model's GOV.UK output accurate — without retraining the model.
Read post - 5 min read
The small models got good — and that changes where government AI can run
Open-weight models now fit the government standard laptop. Here's why that changes what AI government teams can actually run on-device.
Read post - 6 min read
I built a GOV.UK prototyping tool that never goes online
Why Prompt to Page runs entirely on your machine — no API keys, no cloud round-trips, and no draft policy leaving the building.
Read post