Design Exploration: AI Site Design

[shaunandrews]

This design exploration looks at integrating AI into the site creation flow, with a new design process for the agent.

The general flow is:

1. As a brand new user, you've just completed the initial onboard; existing users will have just clicked the + Add Site button at the top of their sidebar. You find you self with a choice for creating a new site:

This isn't a new screen, we've just adjust the first option to mention using AI help you.

2. Selecting the Create a site option takes you to the next step, where you're asked for a site name (the only required field) and have options for providing more context with a prompt, and advanced options if you want to controls versions, etc. The composer here is the same as in our normal chat, just reskinned to feel at home with the WPDS inputs; this means it auto-expands as you type, and you add attachments. It also means any improvements around text editing, links, embeds, etc for the composer also find their way here.

Again, not a new screen, just adjusted to include AI.

3. Once you fill out the inputs, you're taken to the chat and preview screen. (For new users we'll likely want to consider the tour get and getting started flows to cater more to their experience)

Here the agent kicks off a series of skills, including three sub-agents. These sub-agents go off and build html/css/js designs based on the instructions provided by the main session. These designs are showcased in the preview, using an mu-plugin to handle the server side. Users use our tab system to switch between designs, and can annotate and chat with the agent to iterate on any, or all, of the designs.

As the agent edits, it saves iterations for each design, so users can go back and pick up from earlier iterations.

4. Once you happy with once of the designs, you can click Build this design and the agent will mark that design as accepted, and then envoke a skill to convert the HTML site into a proper WordPress block theme. It makes use of skills from across Automattic that I'm still evaluating and tweaking.

--

Its very far from complete, but I want to get conversations going, and to start defining and running some evaluations of models and instructions to ensure quality output.