Hey HN, I'm Yashdeep — I built PrePrompt because I kept writing vague prompts and spending 3-4 turns correcting the model. I wanted something that fixed that automatically without changing my workflow.
The classifier runs pure heuristics in <1ms — no API call. It only fires Haiku when a prompt scores above 38. Simple questions like "what is jwt" get a score of -5 and pass through untouched.
The most interesting part: it learns your stack over time. After 50 prompts it knows you use FastAPI, prefer typed code, etc. and injects that into every optimization without being told.
Happy to answer questions about the MCP hook architecture or the classifier scoring.