I recently thought about combining natural language prompts in code notebooks interface because it seemed like a great use case to me. While working with data, a mixture of low-level capability of writing code and a high-level capability to write transformations or visualization instructions in natural language can be a great advantage. You can code when you want to, and completely skip and offload to an LLM by writing natural language instructions! The notebook executes code and prompts in a single context. Here's something you can try that implements this idea: https://try-slate.com/ Also wrote some thoughts in a post here and a short video demo here.

Nice demo and writeup!

> To save on prompt execution costs, Slate uses a meta-technique in case of array-like datasets to apply transformations

Could you share more about what is happening here?

Yes, sure! Essentially instead of sending the entire dataset to OpenAI APIs and asking the model to apply the prompt transformation, I send the structure of single element in the data set (the type) and ask for a function to apply over the whole dataset to make the requested transformation. The function then is applied in the browser.