AI Expense Report Generator

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

   Create a Next.js 14 App Router PWA for expense tracking. Use Tailwind and Supabase.
   
   Routes:
   - / (marketing + magic-link signup)
   - /app (authed: dashboard with this-month total, category pie chart, recent 10 receipts)
   - /app/add (camera capture + file upload form)
   - /app/receipts (filterable list by date range, category, search)
   - /app/reports (generate PDF for a date range, email or download)
   - /app/settings (categories, default Schedule C mapping, Stripe portal link)
   
   Add next-pwa with a manifest.json and service worker for install-to-homescreen. Mobile-first. Supabase auth with magic links, no passwords. Row Level Security on every table keyed by auth.uid().
   
   Schema: users(id, stripe_customer_id, plan), receipts(id, user_id, image_url, vendor, amount, currency, date, category, notes, confidence, raw_extraction_json), categories(id, user_id, name, schedule_c_code), reports(id, user_id, period_start, period_end, pdf_url, total).

2. 2. Receipt Scanning (AI)

   Implement receipt extraction using GPT-4 Vision with Claude Haiku fallback.
   
   Flow:
   1. User uploads image to Supabase Storage; server action gets signed URL
   2. POST to OpenAI Chat Completions with model gpt-4o, messages [system prompt, user message with image_url]
   3. System prompt: "You extract structured expense data from receipt images. Return JSON matching this schema: {vendor: string, amount: number, currency: ISO4217, date: YYYY-MM-DD, suggested_category: one of [user's category list], line_items: [{description, amount}], confidence: 0-1}. If you cannot read a field return null. Never hallucinate."
   4. Use response_format: {type: "json_schema", json_schema: {name: "Receipt", schema:...}} for structured output
   5. If confidence < 0.7 OR vendor is null, retry with Claude Haiku 3.5 (anthropic SDK, same schema)
   6. Show extracted data inline for user to confirm or edit; save user corrections to receipt row
   7. After user saves, run a small background job that looks at their last 20 corrections and updates a "preferences" JSON on the user row (e.g. "Starbucks near 94110 = Client Entertainment"). Inject preferences into the system prompt on future extractions.
   
   Edge cases: blurry photos (ask for re-upload), foreign currency (convert to user's home currency via Open Exchange Rates API at receipt date), multi-page receipts (accept PDF too, extract page 1).

3. 3. PDF Report Generation

   Build monthly PDF report generation with @react-pdf/renderer.
   
   Inputs: user_id, period_start, period_end
   
   Document structure:
   - Cover page: user name/business, period, total expenses, category breakdown table (category | total | % of spend)
   - Summary page: pie chart of categories (render SVG client-side, capture as image), 12-month trend sparkline, top 10 vendors
   - Detail pages: table of every expense (date, vendor, category, amount, currency), each row has a small thumbnail of the receipt image, sorted by category then date
   - Tax-ready footer: Schedule C mapping table showing category -> IRS line number
   - Appendix: full-resolution receipt images, 4 per page, captioned with the expense details
   
   Trigger modes:
   - Manual: user clicks "Generate Report" in /app/reports
   - Auto: Vercel Cron job on the 1st of every month generates previous month's report, emails via Resend as attachment (use Resend attachments API, max 40MB per email)
   
   Store the generated PDF in Supabase Storage under reports/ bucket with RLS. Return a signed URL valid for 24h. For the free tier, stamp a "Powered by [product]" footer; for Pro, allow user-uploaded logo.