AI agents write UI fast and edge cases badly. This kit gives your agent a solid starting point and tests that catch what agents really get wrong.

What you get

• starter/: a working PWA with no build step: IndexedDB storage, validated JSON export/import, 2 languages, dark mode, service worker with background refresh, manifest and icons

• qa/e2e.js: manifest and icons, persistence, language switch, offline reload and offline writes, dark mode, 320 px overflow, console errors (Playwright with Edge or Chrome)

• qa/axe_scan.js: accessibility scan in light/dark at phone and desktop widths

• qa/monkey.js: random clicks, typing and viewport changes for N steps

• tools/make_icons.js and tools/make_zip.js (forward-slash zip for Netlify-style hosts)

• PLAYBOOK.md: the 7-step loop for each app, a pitfalls table with fixes (null in the DOM, zip backslashes, service-worker caching, axe traps, iOS limits, midnight/time zones, import validation), a quality bar

• ANDROID_NOTES.md: Trusted Web Activity / Google Play requirements and gotchas (12 testers x 14 days, assetlinks, targetSdk)

• A ready skill file for Claude Code that follows the playbook and runs the gates

All scripts pass on the included starter (13 e2e checks, axe clean, 300-step monkey run). Needs Node 18+ and Edge or Chrome. Not verified: macOS/Linux, Firefox/Safari, real phones. Single-user commercial license.