Add an AI policy to contribution guidelines

[jhamman]

AI coding tools are increasingly common in open source development. I'd like us to converge on a shared policy for zarr-python so we can handle AI-assisted contributions consistently — from core devs and outside contributors alike.

The Icechunk project recently adopted a policy that could serve as a starting point:
https://github.com/earth-mover/icechunk/blob/main/icechunk-python/docs/docs/ai-policy.md

The core principle: AI tools are welcome, but you are accountable for every change you submit — you must understand it, be able to explain it, and have fully reviewed it.

Some questions for discussion:

1. Is the Icechunk policy a reasonable starting-point for zarr-python?
2. Should we require disclosure of substantial AI assistance?
3. Where should the policy live — CONTRIBUTING.md, standalone doc, or elsewhere?

I'm personally in favor of embracing these tools with clear guardrails. The primary concern is around maintainer review capacity. Would love to hear other perspectives, especially from @zarr-developers/core-devs.

[d-v-b]

that policy looks good!

[Nyrok]

The accountability principle in the Icechunk policy is the right framing. The problem isn't AI use, it's AI use without structured intent.

When someone prompts "add a feature to zarr-python", there's no constraints block, no reference to your coding conventions, no output format expectations. The LLM optimizes for the literal task and produces code that technically works but doesn't fit the project. Accountability only helps after the fact.

The deeper fix is giving contributors prompts structured around your actual requirements. If the prompt has an explicit constraints block pointing to your contribution guide and a context block with your codebase patterns, the AI output becomes much more likely to meet your standards before review.

I've been building flompt for this: decompose tasks into typed blocks (objective, constraints, output_format) and compile to Claude-optimized XML. Makes it easy to bake project-specific constraints into AI-assisted contributions upfront. Open-source: github.com/Nyrok/flompt

[axtimwalde]

I think this is good. Would suggest to add a paragraph along the lines: PRs that cannot be reviewed in reasonable time with reasonable effort will be closed regardless of their potential usefulness and correctness. Use AI tools not only to develop code and fix issues but to prepare better PRs that can be reviewed more efficiently.

[jni]

I like the policy. More or less mirrors the one we adopted for napari which is based on Zulip's.

I also want to point folks to what I've come to call "Melissa's List" of AI contribution policies. Lots of interesting material there as communities figure this stuff out:

https://github.com/melissawm/open-source-ai-contribution-policies