stylize

Run the same design prompt through different models and compare what comes back.

npm start   # reads ANTHROPIC_API_KEY from .env; http://localhost:4321

Each run starts headless Claude Code (claude -p --dangerously-skip-permissions) in a fresh workspace, runs/<timestamp>-<model>/:

• brief.md: the prompt as sent, chips included.
• content/: the 8 posts (posts.json plus markdown). Images load from R2.
• inspo/: the selected inspo images.
• <direction>/index.html: one folder per direction, written by the agent. A tab appears as soon as the file exists and reloads when it changes.
• meta.json (cost, turns, duration, final notes) and events.jsonl (the agent's full event stream).

Other pieces:

• ref/<n>/: a prompt, its inspo images, and their public URLs (inspo.json).
• app/system.md: harness instructions appended to Claude Code's system prompt (workspace layout, deliverable format, how to screenshot).
• app/screenshot.mjs: the Playwright screenshot script agents use to look at their own pages.
• .agent-config/: an empty Claude Code config dir, so personal settings, memory, plugins and MCP servers stay out of runs.

The agent can run any command. It's told to stay inside its workspace, but nothing enforces that.

R2 bucket stylize (public) holds the post images and inspo. After changing base/ or adding inspo, run npm run upload (or npm run upload inspo).