Fill in a property form, get 3 listing descriptions in 30 seconds instead of 20 minutes.
Built for my dad, who sells real estate. First paid case study in my move into AI engineering.
Next.js 14 · TypeScript · Tailwind · Anthropic API (Claude Sonnet 4.5) · Vercel
- Form collects property details (address, beds/baths/sqft, features, tone)
- API route calls Claude with a cached system prompt and a forced tool for structured output
- Returns 3 variants (Professional, Warm, Luxury), each 100–200 words
- History stored client-side in localStorage, no database
The first version hallucinated. When given sparse input like "Newer build in Bellefontaine, great for kids" it invented a family-friendly neighborhood, nearby schools, and energy efficiency claims — none of which were in the input. In real estate that's a fair-housing liability, not just a typo.
Root cause was a 150-word minimum in the prompt. The model couldn't satisfy both "150 words" and "no inventions" on sparse inputs, and the length rule won. I dropped the minimum to 60, added an explicit "do not invent" list to the prompt, and locked in a regression test using the exact input that broke it.
Full writeup in decisions.md. Project state lives in CLAUDE.md.
v1 live. v2 ideas in backlog.md — waiting on real usage before
picking what to build next.