ENPL
Method
After surgery in 1941, Matisse worked mostly from a bed or a wheelchair. His assistants painted sheets of paper with gouache and pinned the cut shapes to the studio wall where he directed. He did the cutting and the composing himself.
I split my work with AI tools the same way.
Every project gets a short design brief in the repository, in the files coding agents read before they start: AGENTS.md for the rules and a design file for the tokens.
Codex and Antigravity both read AGENTS.md, so neither starts from a blank page and falls back on the average website.
The Atlas engravings began as drawings made by hand in Photoshop, and the generative models only worked from those.
A model is good at variations, but it can’t tell which one is right.
In Antigravity I ask for a plan first and mark it up the way I’d mark up a layout proof. Code comes after the plan is agreed.
In Codex the agent asks before it changes anything outside the project.
Agents take screenshots at phone and desktop widths, in light and dark. I comment on them: Antigravity lets me mark a region of a screenshot, and the browser built into Codex lets me annotate the page itself.
Each comment becomes the brief for the next round.
Matisse’s papers still carry the pinholes of every change, because he kept moving the shapes until they sat right. I keep a decision log for each project: what changed, and why.
Atlas went through 44 rounds before launch, and the log still says why each change was made.
OpenAI Codex
Google Antigravity
Every page was checked at phone and desktop widths in both themes, and the text was checked against common signs of AI writing.
