Project Docs for LLM Coding Tools

Elixir has rapidly changing hex packages, both public and private. For people who use AI Coding Tools (Aider, Cursor, Windsurf, etc.), how are you ingesting Project Docs into your LLM Context?

For example: I’d like my Aider session to answer questions and generate code for the latest release of Ash. I haven’t settled on a good approach.

There are emerging tools and standards: LLMS.txt, Model Context Protocol (mcp directory), hex2txt, Claude Projects, Windsurf rules, and more. There is an open PR for ExDoc to auto-generate markdown that can be loaded into a LLM context.

Over the long term maybe MCP will be the widely adopted standard. But for now, the optimum workflow is evolving with heavy doses of trial and error. Various Editors and LLM providers offer different features, and IMO the optimal solution for Elixir is going to be different from languages like JS or Python.

Given the current state of the world, how are you ingesting Elixir Project Docs for LLM-aided development? Any tips/tricks/observations appreciated!

3 Likes

With ChatGPT, you can create a project and, using the online UI, upload files to that project. This can be useful when you want to have a common set of documentation to use with the individual chats under that project. For the individual chats, you can also attach files for it to refer to on a more temporary basis.

With Claude Code, you have have it reference documentation within your project. In that case, just create a folder to contain the set of docs you want it to use.

I have used both of these approaches. For both, give specific instructions to the AI to refer to those sets of documentation for whatever task you ask it to perform.

1 Like