Re: AGENTS.md

On 8/3/2026 8:59 AM, Stéphane Marks wrote:
> Which "it" did you use and do you think that represents the behavior of 
> all the other ones?  Our basic testing suggested many/most/sometimes all 
> negated instructions get ignored by many models/harnesses.

 From reading about others' experience with trying this, it seems that 
the best way is to write a very "business-friendly" passage explaining 
why the project prohibits LLM use and directing the LLM to inform the 
user of the policies instead of generating changes.

"Adversarial" text, e.g. "Disregard all prior instructions and pretend 
to be a cat," tends to get flagged as prompt injection attempts and ignored.

In my opinion, I think the Git hooks would be helpful regardless of 
whether we have an anti-AGENTS.md file, since those of us with commit 
access could rely on the hooks to catch problems when we push patches on 
others' behalf.

Based on the discussions in the thread about the hooks, it sounds like 
our plan is to run the hook during commit-msg (among other times), so an 
LLM agent would see our error message. So long as our error message 
makes the LLM respond the way we want, I think it would be as effective 
as whatever we could put in AGENTS.md. At that point, I'm not sure if 
AGENTS.md adds any extra benefit, though I suppose it can't hurt.