Case Check
A portfolio review agent that finds the single weakest part of a UX case study's narrative and
says how to fix it. The product contract and evaluation sheet are the source of truth for what it
does and does not do.
Layout
app/ the web page (Next.js): one page plus API routes, no review logic
config/ which Claude model runs each tier; the tunable limits (image cap, thin floor)
content/ every word shown on the page, editable without touching components
src/steps/ the six agent steps, one file each: fetch, classify, plan, checks, verify, report
src/pipeline/ runs the steps in order and streams one event per step
src/lib/ small helpers with no product logic (text clean-up, browser launch, model call, prompt loading)
src/cli/ run the pipeline or a step by hand from the terminal
rubric/ the UXPective storytelling framework as data, editable without touching code
prompts/ every model prompt, one per step, editable without touching code
docs/ decision log
test/ tests, with local HTML fixtures so no internet is needed
Run it
npm install
npx playwright install chromium # once, downloads the headless browser for local use
npm run dev # the page, at http://localhost:3000
npm run fetch -- https://someone.framer.website/case-study-01 # step 1 only
npm run check -- https://someone.framer.website/case-study-01 --target senior # the pipeline, needs ANTHROPIC_API_KEY in .env.local
npm test
Both the page and npm run fetch show exactly what step 1 hands to step 2: the page title, its
visible text with duplicate lines removed, and raw facts (status, word count, image count), or a
failure reason.
Hosting
Deployed on Vercel at casecheck.uxpective.com. The project's Framework Preset must be Next.js.
ANTHROPIC_API_KEY must be set in the project's Environment Variables. On Vercel the fetch step uses the slim Chromium bundled by
@sparticuz/chromium; locally it uses Playwright's own download. See src/lib/browser.ts.
