> On 12 Oct 2026, at 15:56, Richard Kenner <[email protected]> wrote: > >> I rarely use LLMs directly, and have never used them for coding, so I >> don't know, but: is it not possible to feed the AI policy into the LLM >> before you start? It would seem like a flaw if you can't constrain >> the kind of output or interaction that you want, or specify the rules >> that you need to follow. > > Yes, you certainly can and that leads to an interesting idea: why > not have the current policy include instructions that can be passed to > an LLM to constrain what it can do? >
In my experience use of LLMs for production code is more complex than writing a prompt and then taking the resulting patch to the mailing list (that is not allowed for GCC by the current AI policy for non-small changes anyway). It is usually a long session, where you start with an initial prompt, review its output, answer questions it may pose, tell it to rework parts that you find fishy, tell it to do prior research, run extra validation, launch other LLMs to review its work etc. So it would often not be possible to disclose a compact list of instructions given. In my opinion, the cover letter of a patch should already explain the motivation for the patch, the approach taken, major design tradeoffs, and validation done, which actually contains most of the information on how one would guide an LLM to produce the patch itself. If we, as a project, wanted to give some agreed-upon design rules for LLMs to follow we could add an AGENTS.md file to the GCC repo that describes the high-level dos and don’ts, though with the current restrictive AI policy it would probably cause confusion unless its relaxed. >> That is, I'd naively assumed that one of the main advantages of having >> a written policy was that LLMs would be able to work to it. > > I was talking to an architect last week who was complaining that when he > uses an LLM to write contracts, it often comes up with things that violate > his insurance policy. So I suggested that he simply give it the policy > and ask it not to violate that. > > The only problem is that even though the LLM is *capable* of following > that policy, there's no guarantee that it actually *will* and this is > an issue with LLMs in general. But it will a high percentage of the > time and that's probably good enough for this purpose because you'll > know when it didn’t.
