AI/LLM Policy

Luwu contributers are allowed to use LLMs and AI tooling such as Claude Code. Keep in mind that every person (human!) is equally responsible for their changes and commits whether or not a human or a model wrote it.

I will not lie: the Luwu programming language and its core team is greatly assisted by large language models (LLMs) in quickly taking Luau's codebase, iterating on features, and checking for bugs. It has been extremely helpful in fixing the many issues in upstream's type error stringifier in a way that would've been quite difficult or taken forever to do for human authors. But we still have to ensure we use these tools responsibly:

• You must know what your models and subagents are doing while they're working
• You must be able to defend your model's decisions just as you would your own.
• If you or your model made a mistake -- that's fine everyone makes mistakes -- but you gotta fix it.
• You must be willing to fix your LLMs' bad output, reiterate on unoptimial code, fix overly complex or otherwise unreadable code.
• You must be willing to fix its ugly code comments no human would write (inhumanely roundabout wording that puts its most important phrase last, 'x keeps its y', 'x by hand' as examples)

Contributions

• You understand every feature you submit, its edge cases, and you confirm that you've written any RFCs merged into master (almost) completely by yourself without AI assistance in anything other than proofreading and looking for more examples/better prior art/better alternatives/better design, etc. But outcomes of such research should still be written by yourself in RFCs, not by your pet language model.
• You've built and tested the change yourself.
• Commits should not have AI attributions. Co-Authored by Claude is invalid because Claude cannot hold responsibility for a bad commit that breaks user code or causes a security vulnerability. Unless the Claude's a human programmer from France and has been suffering the last 3 years due to his namesake.

Issues and bug reports

• Reproduce a bug yourself before reporting it, and include the code you ran.
• If your AI finds a security problem, you must ensure it actually exists, and if you do you should also write or otherwise produce a REASONABLE, OTHERWISE ACCEPTABLE PR that solves it. But don't spam our issue tracker with a bunch of slop issues or PRs. Keep it to 10 open per author max please. See Reporting Bugs.