Non-Toxic Appliance Verification Platform

AI Prompts to Build This

Copy these build prompts into Claude, Cursor, or your AI coding tool. Create a free account to unlock the full research behind them.

1. 1. Project Setup

   Build the weekend MVP of the "Non-Toxic Appliance Verification Platform": a reviewer uploads a manufacturer's test report, the app extracts the claims, a human approves or rejects each one against its source page, and the public page shows only what was approved.
   
   Stack: Next.js (App Router), TypeScript, Tailwind, Supabase (Postgres, Row Level Security, Storage, Auth with email magic link for reviewers), pdf-parse for page text, the Anthropic API for extraction. Reviewers upload on a manufacturer's behalf this weekend. Deploy on Vercel.
   
   Tables (Row Level Security on; the public reads approved fields only, reviewers read everything):
   - products(id, manufacturer, model, category, slug, status) status is evidence_missing, partially_verified or verified
   - evidence_documents(id, product_id, file_path, uploaded_by, uploaded_at)
   - extractions(id, document_id, raw_json jsonb, model, created_at) the raw model output, kept apart from any decision
   - claim_reviews(id, product_id, document_id, claim_text, source_page, source_quote, status, reviewer_id, decided_at) status is pending, approved, rejected or needs_more_evidence
   
   Screens: /login, /review/[documentId] (each claim beside its source page), /products/[slug] (public), /compare (up to three products), /methodology.
   Env vars (names only): NEXT_PUBLIC_SUPABASE_URL, NEXT_PUBLIC_SUPABASE_ANON_KEY, SUPABASE_SERVICE_ROLE_KEY (server only), ANTHROPIC_API_KEY.
   
   Do not build: billing, membership or manufacturer fees, a manufacturer portal, digital seals, barcode scanning, consumer saved lists, medical or health claims.
   Done when: npm run dev starts, you can sign in, and the four tables exist with Row Level Security on.

2. 2. Core Feature

   Build the one feature that proves the platform: an extracted claim only goes public after a person checks it against its source page.
   1. A reviewer uploads a PDF test report to a private bucket and picks or creates the product. Extract the text page by page.
   2. Send the pages to the Anthropic API with a strict JSON schema { product_model, manufacturer, materials: string[], coatings: string[], food_contact_surfaces: string[], test_standard, test_date, lab_name, claims_supported: [{ claim, page, source_quote }], claims_not_supported: string[], uncertainty_notes: string[] }. Save the raw result to extractions.
   3. Check each claim in code: its source_quote must appear on the stated page. Drop any claim that fails and note how many were dropped. Create a pending claim_reviews row for each claim that remains.
   4. The review page shows each claim next to the source page text with the quote highlighted, and Approve, Reject and Needs more evidence buttons. Save the reviewer and the time.
   5. The public product page shows only approved claims, the evidence date, the test standard and lab, and one status: verified when every key claim is approved, partially verified when some are, evidence missing otherwise. Each status links to /methodology.
   6. /compare shows up to three products side by side with their statuses and approved claims.
   Rules: never write "risk-free" or any medical claim. Rejected and pending claims never appear publicly. Show "Unknown" instead of guessing a missing value.
   Empty state: with no products, show the upload form for reviewers and an empty public list.
   Done when: a sample report produces claims whose quotes all appear on their pages, an invented quote is dropped, a pending claim is invisible on the public page, approving a claim publishes it with its date, and the public page never uses the word risk-free.

3. 3. Landing Page

   Build a one-page landing site for the Non-Toxic Appliance Verification Platform, an evidence database for people choosing appliances.
   Hero: "See what the evidence says about the appliances in your kitchen." Sub: "Every claim shows its source, its test and its date, and says what is still unknown." One button: Join the waitlist.
   Sections: a sample product page with its three possible states (verified, partially verified, evidence missing), how it works in four steps (upload, extract, human review, publish), why a person checks every claim, and an FAQ on what verified means, why we never say risk-free and how brands can submit a report.
   Waitlist: store the email in a waitlist table in Supabase. No other service.
   Style: Geist, a warm white background, ink type, one calm green accent.
   Done when: the page renders on a phone and a submitted email appears in the waitlist table.

4. 4. Branding Package

   Use a design or image tool for this one. A coding agent cannot draw a logo.
   Brand for the Non-Toxic Appliance Verification Platform: pick a short working name, then design a wordmark and an icon that suggest a cup outline with a check inside it.
   Colors: warm white, ink and one calm green. Type: Geist.
   Deliverables: wordmark, icon, three status labels (verified, partially verified, evidence missing) that differ by shape and word as well as color, an evidence line style (source, test, date) and one launch graphic.
   Done when: each deliverable is saved in one folder and the three status labels are distinguishable without color.