What's the best way to create llm consumable documentation for hyperdiv?

[idvorkin]

I love hyperdiv, and want LLMs helping me write them! Hyperdiv has awesome docs, so thinking about how to get the docs to the LLM!

FastHTML tackles this by publishing llm.txt, documentation to be used by the LLM. Is there a recommended path to do this from the docs app?

LLM Text in general: https://llmstxt.org/
FastHTML's version w/links : https://docs.fastht.ml/llms.txt
FastHTML's version w/links inlined: https://docs.fastht.ml/llms-ctx-full.txt

[idvorkin]

OK, I went a bit crazy and extracted them using ast-grep. I'm certain there's a better way, but this was fun, and comes close:

https://github.com/idvorkin/LinqPadSnippets/blob/0b1efd0cafc082644801a0dc77dd90c7507f43b5/python/dump_hyperdiv_docs.py?plain=1#L17

If we want to merge in some variant of this holler.

[idvorkin]

Here are the generated docs:
docs.md

[idvorkin]

And if you want a giggle, I used GPT to figure out the document sort order :)

[mariusnita]

How did this experiment turn out?

My idea that I haven't gotten around to was to build a tool that generates Hyperdiv code, by feeding granular component-level documentation and function signatures to Claude/GPT API, probably using function calling.

So you can be like "add 2 side by side buttons. When I click the one on the right a dialog opens with these contents: ..."

And Claude/GPT will then request the detailed documentation and call signatures for "box", "button", "dialog", etc. to generate precise code.

Initially it would be fed some kind of overview (probably like what you did) and a total list of components that it can request docs for.

I agree that a basic 'hyperdiv-llm.txt' would be useful to provide though, though I think the docs for the full API would generally be too large to include in a prompt.

[idvorkin]

I haven't used this in a while, but I think it just worked. It's only 21K tokens, which isn't bad (and even less so now that we have caching). I think future tools will use RAG over the docs (I suspect cursor does this) to cut down the context window furthor.