Skip to content

Designing for an AI host.

We designed the editor, the AI interaction framework and the template library for a builder of apps that run inside ChatGPT and Claude.

Layo AI design workflow and mobile canvas
Client
Layo
What they build
Builder for apps inside ChatGPT and Claude
Our work
Product design, AI interaction framework, Templates
Timeline
December 2025 to March 2026

The company

Layo lets anyone design, connect data to, publish and measure apps that live inside ChatGPT and Claude. We joined a month after it was founded.

The problem

Interactivity came first: there was no way to add button actions or modal triggers. The founder wanted people to feel they could build an app in five minutes. Later came the blank-screen problem, and how chat should live inside a visual builder.

Templates beat a perfect blank prompt for first-run value.

The work, piece by piece.

01

Interactions

A side panel instead of controls on every component, drag-to-connect, and play-mode prototyping.

02

Data binding

Color-coded states for connected, hard-coded and broken, and a fix connection button that hands the error to the AI.

03

The AI assistant

Select a component and it becomes the chat's target. The canvas locks while the agent works, with stop where send was.

04

An AI interaction framework

Cold start, plan, execute and chat-only modes, plans written as documents, visible memory, diffs, and confirmation before destructive actions.

05

A template library

Booking, analytics, coaching, project management and more, plus a public app store page.

The AI editorThe Layo editor with the AI chat panel explaining how it built a hotel search widget, previewed as a ChatGPT app

How it went

The key call: templates over perfecting the open-ended chat. They deliver first-run value faster and more reliably.

What other teams can learn from it.

  • Templates instead of a blank prompt
  • Stop, queue and interrupt while an agent runs
  • Visible reasoning and tool calls
Assert's review layout: a dark split diff with a file tree, per-file review status and the clusters barNext · Assert LabsAgents write more code. Humans need a better review.

Building something like this?

Start a project