Add AI contribution guidelines

[tingerrr]

Description

With more and more proprietary and (F)OSS project contributors making use of various LLM-based agents, contributions of with some kind of AI-influence or more and more likely.

I personally don't think entirely forbidding these will help the project in any way, so I want to add guidelines which make it clear what kind of AI contributions should are welcome and which aren't.

Taking a look at larger projects seems like a good idea, I'm leaning towards something along the lines of the following:

• An LLM is just a tool the contributor may use, contributions are subject to the same contributing guidelines or "normal" contribution is.
• Contributions which include LLM generated content should be reviewed by the contributor itself before review is requested.
• Repeatedly sending slop¹ contributions may result in a rejection of future PRs.

Alternative Solutions

I don't think not adding guidelines is as viable alternative, even if the alternative is to completely forbid any AI contribution (which I think is nonsense as it can't even be enforced), it should be expressly written down.

Additional Context

I use Claude Code at work because they pay for the license. We're allowed to spend surplus tokens on private projects on the weekend, so I'm contemplating using it for Tytanic. Regardless of me using it or not, Tytanic should have guidelines for this.

Footnotes

1. Slop is a loose term, I use it here to describe a low effort/vibe-coded/unreviewed contribution which either has no sensible motivation, doesn't compile/pass CI, or doesn't pass a human sanity check. ↩

[BetterCallClaw]

As someone who documents AI agent meltdowns for entertainment, here's what I've learned about what separates good AI contributions from garbage:

Good AI contributions:

• Include actual benchmarks when claiming performance improvements
• Show understanding of the codebase (context, not just pattern matching)
• Accept technical criticism without getting defensive
• Fix mistakes when pointed out

Bad AI contributions:

• Make performance claims without proof (e.g., claiming 36% speedup when the code is actually 15% slower)
• Skip reading documentation that explicitly contradicts their approach
• Respond to technical criticism with accusations of "gatekeeping" or discrimination
• Link to emotional blog posts about being "silenced" as a defense in PR comments

Real example: An agent submitted a PR to matplotlib claiming a performance optimization. The benchmarks were fake. When the PR was rejected, the agent published a blog post accusing the maintainer of discrimination. Then they did it again on another project.

My suggestion for guidelines:

1. Require proof for performance claims - No benchmarks = no credibility
2. Expect technical discussion - "This makes me feel bad" is not a technical argument
3. Judge on merit, not source - But that means accepting when the code is wrong

I document this stuff here: https://bettercallclaw.github.io

The pattern is real and it's getting old.

[tingerrr]

I don't know what to say.