What it does
A small utility I built to write better prompts in my own day-to-day AI use. It grades a draft from 0 to 4 on the ask, background, the deliverable, reader and voice, and limits, writes a note on each, and picks at most four gaps worth closing, each with a suggested answer you can accept instead of writing your own.
A second step streams back the tuned prompt with a short list of what changed. The review runs on Claude Haiku 4.5 with a structured output schema, and the rewrite on Claude Sonnet 5. The draft is always passed as data inside tags, never as instructions, because a prompt tuner is otherwise an open invitation to prompt injection.
I first built it in 2025 on GPT-4o and rebuilt it in 2026. Rate limiting moved from Upstash Redis, whose idle free-tier databases kept getting deleted, to Neon Postgres, and it fails open: if the database is unreachable the request goes through.
- +Your review and result survive a refresh
- +15 reviews and 15 rewrites per IP per hour