5 April 2026

Using AI skills to make complex dev environment accessible to designers

The real unlock wasn’t “cheaper code.” It was turning team conventions into AI skills, so a designer with no engineering background can work in a production-grade environment without constantly falling into traps.

Main point: The real unlock for designers using AI isn't cheaper code. It's skills: our team's conventions written down so the AI follows them every time.

Why plain AI help falls short

Give a designer Cursor and hope, and you get code that compiles but ignores your conventions, misuses components and invents imports. The AI knows React. It doesn't know your repo, for example:- use semantic Tailwind classes (bg-background, text-foreground) backed by tokens in src/os/design-system/tokens/, not one-off hex values- keep OS state in the Zustand store, not ad hoc localStorage- put new apps in src/apps/<app-id>/ and register them (AppID, WindowShell, src/os/config/apps.tsx)

What a skill is

A short markdown runbook for one task, triggered by a phrase. Good skills are specific to your system, spell out the don'ts (like wrapping Radix ScrollArea in a flex child with min-h-0), and are maintained like code.

The skills that matter most

  1. First-time setup: from zero to a running app in one conversation, with no decisions left for the designer
  2. Create a new app: the full structure, wired in and running
  3. Audit my app: checks tokens, accessibility, imports and conventions before a PR, and explains the fixes
  4. Start a new branch: plus a summary of what changed on main

Skill vs prompt

A prompt gets the AI's best guess. A skill gets it following a procedure written by someone who knows the repo. It's the difference between "build me a kitchen" and handing over the drawings. And skills compound: write once, and everyone gets the same guardrails.

For teams

Don't fixate on what AI can do. Focus on what it will do consistently in your system. The AI doesn't make the system accessible. The skills do.