PuxaPapo
Reads the chat screenshot and gives you what to reply — and why it works.
[ The project ]
Multimodal AI conversation assistant. You send a chat screenshot, pick a tone and a goal, and get the context read back, the interest and coldness signals, and suggested replies. Six tools over one engine, passwordless signup and a recurring Pix or card subscription.
[ stack ]
- Next.js 16
- React 19
- TypeScript
- Google Gemini
- Saída estruturada
- Supabase
- Postgres + RLS
- AbacatePay
- Vercel
[ Technical decisions ]
- 01
Six tools over one engine: analyze a screenshot, write an opener from scratch, improve a draft, refine a suggestion, a practice mode with a conversation simulator, and a weekly coach report.
- 02
Gemini's output is declared as a schema rather than asked for in prose: context, green signals, red signals, a basic suggestion and the premium ones, each with a micro-explanation of why it works.
- 03
A paywall you can't defeat with devtools: premium suggestions are truncated on the server and the full text never reaches the client. The blur on the front is purely cosmetic — the protection sits before the response leaves.
- 04
The other person's persona accumulates per conversation from previous screenshots, recomputed on each new analysis, with multi-turn history capped and sanitized before entering the prompt.
- 05
User taste learned from 👍/👎 feedback: recent liked and rejected examples feed back into the prompt and calibrate the style of what comes next.
- 06
Profile (age, what they're looking for, vibe) and conversation goal change the strategy of the reply, not just its tone.
- 07
Per-user rate limiting, image validation and a cap on images per analysis — the expensive path is the one that calls the model, so that's the one that needs a gatekeeper.