AI products are changing what frontend work means, but not in the way the headlines suggest.
The models got good fast. A capable model can now write, summarize, plan, call tools, and hold a conversation. What did not follow automatically is the part people assumed would: a good model did not turn into a good product on its own.
That gap is the subject of this site.
A working model is not a working product
What a person actually uses is not the model. It is the interface around it: how the model’s output is rendered, what actions it can take, and where a human stays in control of the decisions that matter.
A model that drafts a contract is impressive. A product that shows the draft, marks what changed, and asks before it sends anything is useful. The distance between those two is interface work, and most of it is frontend work.
Chat is usually the first surface teams build, because a text box is the easiest place to accept messy human intent. It is a good entry point. It is rarely the whole answer. The strongest AI products pair conversation with real components, structured data, explicit actions, and clear trust boundaries.
Three layers
I think about AI interfaces as three layers that depend on each other.
flowchart LR A["Conversation<br/>captures intent"] --> B["Structured interface<br/>makes it inspectable"] B --> C["Trust & action<br/>keeps people in control"] C -->|"feedback"| A
- Conversation captures intent. A person says what they want in their own words.
- Structured interface makes the result inspectable. Cards, tables, forms, diffs, and previews turn model output into something you can scan, compare, and edit.
- Trust and action keeps the person in control. The system is explicit about what is a suggestion, what is a draft, and what will change the world.
Most AI features that fail skip one of these. A product with only the first layer is a chat box that makes the user do all the reading. A product that jumps to the third without the second asks people to approve things they cannot really see.
Why this matters now
Raw model quality is becoming a commodity. Several providers ship capable models, and the best one this quarter is not always the best next quarter. When the model is no longer the differentiator, the product is. That product lives in the interface layer, which is where a team’s own judgment about clarity, trust, and workflow actually shows up.
What this site is
Interface Lab is field notes on that layer: chat UX, generative UI, structured interactions, trusted rendering, and the protocols starting to standardize how agents talk to frontends. It is written for the people who build these products: frontend leads, senior engineers, and the CTOs and founders deciding what to build.
I am not interested in the claim that chat is dead, or that frontend is dead, or that the model will generate every screen by itself. None of that matches what ships. The interesting work is more specific: how do you wrap a probabilistic model in an interface people can trust with real decisions.
Where to start
If you want a path through the core ideas, read the four-post series AI Interface Foundations. The posts build on each other.
- Chat is a surface, not the product: why collapsing everything into messages breaks down, and how to route model output to the right surface.
- Structured outputs need structured interfaces: connecting model schemas to components so the UI stays predictable.
- Trust boundaries in AI interfaces: separating suggestions, drafts, and actions so people stay in control.
- From prompt box to workflow: giving AI work durable state so it survives more than one message.
Each one starts with a real problem and ends with something you can build.