How I work
Anyone can get a screen out of an agent. Whether the hundredth one still matches the first is a foundation problem, not a prompting problem.
01
Foundation first
Tokens, voice and prior decisions — encoded where the AI can read them.
02
Rules in the codebase
Not a wiki nobody opens. So the hundredth screen matches the first.
03
The system polices itself
Drift is flagged, named and located automatically. Nobody audits by hand.
04
Prototype against the real thing
The live codebase, not a picture of it. Tested to be corrected, not approved.
05
Seed it with real data
Lorem ipsum gets a polite reaction. Real data gets an argument.
06
Judgement stays human
Review gates, do-not-build lists, taste. Nothing ships that I wouldn't sign.
Design for the AI era
AI raises the ceiling.
People still set the bar.
AI assists. Humans decide.
AI’s job is to elevate the people doing the work — sharper research, faster iteration, more options on the table. The taste, the decisions, and the accountability stay human.
Know when not to use it.
The skill isn’t using AI everywhere — it’s knowing where it pays off and where it degrades the work. Some problems want a model. Some want a customer conversation, a whiteboard, or a hard no.
Built into your team, not around it.
AI lands inside the structure you already have — your engineers, your codebase, your design system, your review process. Briefed against the real infrastructure, shipped through the process your team already trusts.
01
It’s all about context
Everyone talks about design.md files. A markdown file is just a text file — the format was never the point. What matters is what you encode in it: your design tokens and spacing logic, your brand voice, your component decisions and, critically, the reasoning behind them. Markdown happens to be a format both humans and AI read fluently, which makes it the natural vehicle for any kind of context: design principles, engineering constraints, tone of voice, things deliberately not built and why.
Context
What the AI knows. Design tokens, component inventories, brand voice, prior decisions and their rationale. The richer the context, the less the AI invents.
Rules
What it must respect. Hard constraints that live with the work: use these tokens, never these values; extend these components, don’t fork them.
Skills
What it can execute. Codified, repeatable playbooks — analysis, story breakdown, applying the style guide. Written once, run by anyone.
The tools will keep changing. The principle won’t: the quality of AI output is a function of the quality of its context.
02
Design systems the AI can read
An AI-ready style guide isn’t a PDF — it’s data an agent can read. Design tokens define the values; a component library defines the patterns; Storybook documents behaviour; Figma holds design intent, now readable by agents through MCP. Connected, the style guide stops being documentation that drifts out of date and becomes the live contract every generated screen is checked against.
A new generation of tools makes that contract visible before anything ships. Pencil.dev treats designs as files living in your git repository, editable in the IDE, with agents reading and writing them through MCP. Paper renders a free-form canvas in real CSS, so what you design is already the code you’ll ship. I evaluate these hands-on and critically — some earn a permanent place in the workflow, others get walked away from. The evaluation discipline matters as much as the adoption.
03
The real trick: rules in the codebase
Context that lives in a wiki gets ignored. The unlock is putting the rules where the AI actually works — in the repository. Rules files sit beside the code. Tokens are imported, never re-invented. Component usage is constrained by the system itself. Every agent session starts with the same guardrails, so the hundredth AI-built screen matches the first.
That is what makes the practice safe to scale beyond designers: anyone on the team can direct an AI to build a feature with the security and confidence that it will look and behave the way it should. Design stops being a bottleneck and becomes the operating system the whole team builds within.
04
Where the human stays
None of this removes judgment — it concentrates it. Taste, product strategy, what not to build, when the AI’s answer is confidently wrong: that is the work. I keep explicit review gates ahead of anything customer-facing, structured handoffs with acceptance criteria and do-not-build lists, and a simple standard: no AI output ships that I wouldn’t put my name on.
05
Where this is proven
At Chime Labs, where I work as a full-stack product designer, we build AI with AI — the style guide is encoded as rules, playbooks run as skills, and AI-augmented engineering gives a small team the leverage of a much larger one. At Smokeball and SkillsAware, I led teams through the early shift from static Figma mockups to AI-assisted prototyping — practices that have compounded as the tools matured.
And this site is its own demonstration: designed, restyled, and maintained by an AI agent operating under exactly the rules and context described above.
Working on scaling design with AI — or deciding what to trust it with? Talk to me.
Tell me what you’re building. I’ll tell you if I can help.
Email me →