Add Nix agents.md proposal.

[crertel]

No description provided.

[trueNAHO]

AFAICT, the general FOSS consensus, including the NixOS ecosystem according to NixOS/nixpkgs#410741, seems to be to ban LLM contributions or disengage with LLM content.

Before making LLM-generated content easier, it should be established whether LLM content is even welcomed, which does not currently seem to be the case.

[trueNAHO]

It sounds like the real problem is about improving the human documentation in general, instead of adding incompatible non-human documentation.

This seems to be a subset of NixOS/nixpkgs#387072, which is very likely a lot more effort than just "medium".

[crertel]

We've had years to improve the human documentation--if we can backdoor in incremental improvements there for people that are excited about AI, that's a win.

[crertel]

The thing is, people are already using LLMs to make working with Nix and NixOS easier.

That genie isn't going back in the bottle. So, here's our chance to make things easier for everyone to use Nix and NixOS correctly when partnering with LLMs.

Whether or not we accept those contributions upstream is orthogonal to how we help somebody trying to deploy their own machines or check their config for better ways of doing things.

[Eveeifyeve]

I would say this needs to be benchmarked before going ahead and adding this. Three question's are to be answered does this drastically bring down the performance, if so where is the performance is decreasing and what could you do about it.

I would like to mention there has been done research recently that some Agent.md files actually perform worse being there than not: https://arxiv.org/abs/2602.11988, So to avoid this I would say stuff that is "needed" (It can also towards the model in the right way) and not in codebase needs to be added only. If it doesn't meet those criteria then it shouldn't be added to the Agent.md file.

At the end the conclusion, this needs to be benchmarked before adding it.

[crertel]

@Eveeifyeve

So, this is less benchmarking (see my other proposed project) in the quantitative sense and more containing a go/no-go check:

• Demonstrate improved code generation and output for an LLM of their choice when using this file

The purpose of this agents.md would be to provide guidance to the LLM around best-practices:

• When writing a flake, are we using modern practices (home-rolled vs flake-utils vs something else)?
• When working on a nixpkg, are we including good tests using VMs, are we including update scripts?
• When doing anything, are we running nixfmt after?

These are the sorts of things that the agents.md would help with (and I'm sure there's other tribal knowledge that would be worth encoding and keeping current). The "benchmark" then is exactly as mentioned in the deliverables--showing that it either helped or hurt compared to that stock LLM.