Add a GEMINI.md file with instructions for the CLI

[mgeisler]

I’ve been playing with the gemini CLI tool
recently (https://google-gemini.github.io/gemini-cli/) and learned that it will read instructions from it’s own kind of README.

So I asked it to generate one and use it as a basis for this file. I took the liberty to expand the style guide a little at the same time to incorporate that instructions I gave Gemini while generating the many language fixing PRs.

[fw-immunant]

Honestly, I would really rather not encourage LLM use in this repository. The point of teaching is to communicate semantic meaning, which is an activity performed between beings who have access to semantic meaning. Putting a language approximator in the middle is just encouraging handwaving rather than conveying precise understanding.

[djmitche]

100x agree with @fw-immunant. I realize I'm swimming against the current industry fad, but it seems particularly important in this repo.

[gribozavr]

I would like to reassure you that we're not lowering the code review standards.

[mgeisler]

Hey folks! I'll admit that I'm sceptical of LLMs myself, but I've been having a lot of fun using them lately for non-code tasks 😄

Basically, I've found them useful for tasks with words where I don't already know the answer or where there isn't a strictly "right" answer. When trying to do coding tasks, I find them infuriating: Gemini acts like a very determined intern who wants to solve the task, with no regard for how. It's like working with a poor developer who has no taste. Very weird experience and I feel that we'll all have job security for years to come 😉

However, back to text: we now have about 400 pages and I don't have the skills or time to make the grammar consistent across that many pages. So I wanted to see what happens if a LLM reads the whole thing and does it's best. I would be happy if you could look at the string of PRs merged over the weekend to see how good or bad a job it did.

Related, I put up #2899 and 2898 earlier tonight, generated 98% by Gemini. I'll very much agree that they're a bit bland, but there might be bits and pieces that are helpful in there? Feel free to tell me otherwise! Note that #2900 is hand-written, no LLM there!

[mgeisler]

For a PR like #2809, I found Gemini really cool: it enables me to evaluate something I could never do in a reasonable time before.

For this, the gold standard (and perhaps a great project for @michael-kerscher, @egithinji, or some other 20% contributor?) would be to integrate https://github.com/google-github-actions/run-gemini-cli/.

I would love to be able to write @gemini-cli /review in a comment and get an evaluation a few minutes later for all the translation PRs. If there is a community of interested reviewers, then by all means, they should review things by hand! But we have a few languages where there is just one person sending out PRs, and there I think it would be a great and positive experience to receive a quick robotic review.

This would be a cool way for Google to showcase its AI, hint, hint 😉