Your design system is a suggestion to an AI agent.
It reads your tokens and picks whatever feels close enough. It generates 10 screens with 10 different spacing values. None of it is wrong enough to notice at a glance. All of it adds up.
The fix is not better prompts. It is a 6-layer constraint framework.
@stephenhaney Yes, but HTML is the language for visual communication between humans and agents, not necessarily the best source of truth for agents themselves.
The moment HTML becomes canonical, complexity, navigation, hidden dependencies, and ambiguity start creeping in very fast.
I think HTML makes sense for humans, but for agents it opens a Pandora’s box of complexity.
Links, navigation, hidden dependencies, interaction logic… suddenly the agent is exploring instead of reasoning through a constrained system.
And the moment we add rules to make HTML reliable, we slowly rebuild a Markdown-like environment again.
The future probably isn’t Markdown vs HTML, but constrained semantic systems optimized for agent reasoning.
Your design system is a suggestion to an AI agent.
It reads your tokens and picks whatever feels close enough. It generates 10 screens with 10 different spacing values. None of it is wrong enough to notice at a glance. All of it adds up.
The fix is not better prompts. It is a 6-layer constraint framework.
@hamen omg this is amazing, testing it asap
one thing I think is missing: `useWhen` annotations
it tells the agent what each component is, not when to pick it over an alternative
that's the gap between human-readable docs and agent-optimized ones
Agents don't fail because they're not smart enough. They fail because you gave them the wrong shape of information. Constraint over instruction. That's the whole thesis.
@tirumalab_ch @sanderv@stephenhaney@stevelauda_@figma@paper It really has to do with @paper being fully built to be the AI-first and Figma is trying to evolve into that but there is too much on it to be fixed to be able reach that
@tirumalab_ch @sanderv@stephenhaney@stevelauda_@figma@paper It could be if you are asking to use pre-existing components without clear context or the file structure, or even the way the tokens are structured 🫠
Most people are optimizing for token count. Wrong metric. Precision per token is what matters. One well-structured constraint beats a paragraph of explanation every time.
Betting everything on a single markdown file (design.md / skill.md) will let you down. Real quality, consistency, and efficiency require a proper system — not one file.