guidance-ai/guidance ? reverse-engineered prompt

Build me a Python library for steering large language models with simple, readable code. I want to be able to write prompts in a Pythonic way, switch between system, user, and assistant messages, and capture generated text into variables. It should also let me constrain outputs with things like regular expressions, picking from a list, and more structured grammar style generation, so the model can only return valid text for my use case.

Please make it work with common model backends and feel smooth in Jupyter notebooks too, with a nice interactive widget if possible. I’d also like a way to define my own reusable guidance functions, so I can compose prompt logic instead of rewriting it every time. Include a few simple examples and make it easy to run locally for testing without always calling a real model. If you need to look up current docs online, feel free to do that.