I have written system prompts myself and mostly they are always a free text, a block of text based on which AI projects function. If someone makes a bad edit to it, production reflects that. And we have no way of finding it but to read the system prompt and understand the semantic impact of that change ourself before fixing it. But what if that system prompt was more like a code? Could be validated, tested and then deployed? That's what I have built using claude code, there are commands and agents to convert a free text prompt into a typescript schema, with proper contracts. We can run diff between 2 or multiple prompt versions to see how the behaviour has changed. We can test this prompt using adversarial test cases designed to break it (what claude will generate), or even run it on our own ground truth data. How to run it? All you need is claude code and this project on your system. In examples you can find 3 samples. The netflix one is my favourite.

Promptx

A type system for LLM prompts. Ship with confidence.

A type system for LLM prompts. Ship with confidence.