prompt-to-paint

How long a coding agent takes to put something on screen that a human can react to, measured by screenshotting the page every second and scoring each frame against the brief.

Skeleton first: telling agents how to render early

8 October 2026 · ops-dashboard brief · 24 runs · four agent/model pairs, told vs told how

Told to save a bare skeleton first, then one section per save, 12 of 12 runs built the page in four or five visible stages. Told only to render early, 0 of 12 did.
First render came 1.4–2.6x sooner for every agent, down to 7.5s for Sonnet 5.5 in Claude Code.
The finished page came about twice as late, so area under the curve improved for one agent of four.

Replay every run · Race chart

Dense ops dashboard: five agent/model pairs, three vendors

7 October 2026 · ops-dashboard brief · 15 runs · every run replayed side by side

15 of 15 runs went from a blank page to the finished dashboard in one step, despite being told to render early.
Every run scored 1.00, so first render alone decided the ranking: 18.9s (Sonnet 5.5 in Claude Code) to 38.6s (Gemini 3.8 Flash in Pi), medians.
Judges from OpenAI and Google re-scored every run; 29 of 30 re-scorings matched the original exactly.

Replay every run · Race chart

Task board on a running Vite + React project

7 October 2026 · todo-app-scaffolded brief · 15 runs · first and later changes judged

The harness ran the toolchain and every save hot-reloaded, yet 13 of 15 runs still went from blank to finished in one step.
One GPT-5.5 run built in stages: unstyled at 12.5s, styled at 23.5s, finished at 27.1s.
Of nine changes after a first render, three were judged improvements. On the dashboard, none of eight was.

Replay every run · Race chart