AI drafts, I decide
AI is great at options and first drafts. Choosing what's right for real users is still my job.
AI Lab
Words in. Interfaces out.
A prompt is just words, and words are where every interface starts. Here's how I use AI workflows and tools to go from words to design to code, and the rules I keep so the result still feels human.
❯ pyrus run --from words --to code
✓ research AI synthesised 12 interview notes → 4 themes
✓ ideate generated 10 user flows → kept the best 2
✓ design 3 layout variants in Figma → 1 refined by hand
✓ build <CheckoutStep /> + tests, reviewed line by line
▸ test waiting for real humans to try it… 👀
# AI drafts, I decide. Done in days, not weeks.
// rules.md
My rules
Real people over predictions
AI can summarise research, but it can't replace watching a real person use the interface.
Own every line
Generated code gets reviewed, tested and made accessible. If I ship it, I understand it.
// workflow.yml
The workflow
Understand the words
AI helps with
Summarising interview notes, clustering feedback into themes and drafting research questions.
I own
Talking to users, spotting what they don't say, and deciding which problem is worth solving.
- ChatGPT
- Gemini
- Perplexity
- Notion AI
Words into flows
AI helps with
Generating user flows, edge cases and "what if" variations I might not think of alone.
I own
Sketching the flows that fit the product's goals and cutting the ones that don't.
- ChatGPT
- FigJam AI
- Whimsical AI
Flows into design
AI helps with
Placeholder copy, image concepts, layout variations and checking colour contrast.
I own
Visual hierarchy, the design system, and the details that make it feel crafted.
- Figma AI
- Midjourney
- Figma
- Stark
Design into code
AI helps with
Scaffolding components, writing tests, refactoring and generating first-pass UI from a prompt.
I own
Architecture, accessibility, performance and the motion that makes it feel alive.
- Cursor
- GitHub Copilot
- v0
Code into confidence
AI helps with
Drafting test scripts, summarising session recordings and flagging accessibility issues.
I own
Running sessions with real users and deciding what to change next.
- Maze AI
- Playwright
- axe DevTools
- Lighthouse
// ai-toolbox.json
Toolbox
Think & research
- ChatGPT writing, reasoning, synthesis
- Gemini brainstorming
- Perplexity sourced research
Design
- Figma UI & systems
- Midjourney visual concepts
- Stark accessibility checks
- Framer Motion UI animation
Build
- GitHub Copilot pair programming
- Cursor AI editor
- v0 UI from prompts
- React components
- Next.js apps & sites
- TypeScript typed UI code
- Tailwind utility CSS
- Storybook component docs
- Node.js tooling & APIs
- Vercel deploys
Test
- Maze usability testing
- Playwright end-to-end tests
- Lighthouse performance
- WCAG accessibility standard
Automate
- n8n workflow automation
- Zapier quick glue
- MCP servers tools for AI agents
Exploring next
- Generative UI interfaces built on the fly
- Voice interfaces talking to products
- AI agents multi-step workflows
From the blog
Read the full breakdown.
I write about the AI workflows that work for me, the ones that didn't, and what they mean for UX.
My AI-powered design-to-code workflow- 01 Research
- 02 Ideate
- 03 Design
- 04 Build
- 05 Test
