Jan Collaert I, after Jan van der Straet (Stradanus), The Invention of Book Printing, plate 4 of Nova Reperta, ca. 1600. Engraving. The Metropolitan Museum of Art, 34.30(5). Public domain, Met Open Access (CC0).

Let the model write the screen. Own the catalog.

Good morning.

This week the model started writing the screen, the workflow, and the watch list. The products that work are the ones that fence it in.

On the cover, compositors set any page you like from the same case of type. The shop owns the letters, the press, and the proof. That is the week.

1. Generated UI is a compiler, not a canvas.

A teardown of ChatGPT’s Intelligent UI shows the model never writes raw front-end code. It writes a small language that mixes prose, tags for approved components, and a little state. Every few hundred milliseconds a server recompiles it, repairs half-written pieces, and drops anything invalid. A sandbox with no network access turns the result into UI operations, and the app draws them with its own native components. An iframe is the escape hatch for the rare thing the catalog can’t do.

rabi_guha · long version

Same week, Claude artifacts cost half the usual usage on paid plans through Oct 15, with a design system captured once from screenshots or a URL, connector-fed dashboards, and a quiet change: a shared artifact runs on the viewer’s usage, not the author’s. An agent-written new tab rebuilds your homepage every day. A sales CRM built in about two days used one conventions file as its whole spec.

cinkotweets · Thariq · Marcel Kargul

Safe, on-brand generated UI comes from a fixed catalog and a compile step, not from letting the model write HTML.

2. Skills are the app layer. The harness is the enforcer.

Josh Rosen looked inside the big skill bundles from OpenAI, Stripe, Microsoft and others. The good ones are organized around units of real work, put a routing skill up front, and keep tools behind connectors. It’s a service architecture written in Markdown. But handoffs are untyped, approvals aren’t enforced unless the harness does it, and runs aren’t reproducible. Reviewable HTML plans and automatic handoff notes are small moves the same way: make agent work readable and resumable.

Josh Rosen · html-plan · auto-handoff

3. Attention, not generation, is the bottleneck.

One engineer wants a software foreman, not a factory: an agent that learns what normal looks like in production, investigates anomalies, and comes back with findings or a small fix. A performance-watchdog prompt meant to run for a year does exactly that. Guided reviews explain a pull request core change first. And the honest cost: agent waits are frying our dopamine.

fractaledmind · Rhys Sullivan · Capy · swhotch

Try this

1. List the five components your users actually need before you let a model generate any UI.
2. Move one required approval out of a prompt and into code.
3. Point one agent at a stream of production data and ask it what looks wrong.

Let the model write the screen. Keep the catalog, the harness, and the review. Coffee’s on.