Re: AGENTS.md and CLAUDE.md Are Concessions to Proprietary Services, Not Defenses Against Them
Jean Louis <[email protected]> Tue, 04 Aug 2026 12:31:02 +0300
| Newsgroups | gmane.emacs.devel |
|---|---|
| Message-ID | <[email protected]> |
On August 4, 2026 10:09:41 AM GMT+03:00, Yuri Khan <yuri=2Ev=2Ekhan@gmail= =2Ecom> wrote: >On Tue, Aug 4, 2026 at 1:36=E2=80=AFPM Jean Louis <bugs@gnu=2Esupport> wr= ote: > >> This is not FUD=2E It's a principled question: why should GNU adopt a >> convention created by a proprietary company > >That=E2=80=99s what GNU is, historically: A reimplementation of conventio= ns >created by a proprietary company=2E (UNIX) Creating free software system based on some principles found in proprietar= y software is totally different to adopting vendor specific file format tha= t evidently cannot protect what we strive to protect=2E >> Let me take example of OpenCode; >> >> Ignore AGENTS=2Emd and and improve my-package=2Eel > >You=E2=80=99re talking like you assume the LLM to actively look for and r= ead >that file=2E That is not what happens=2E No, I don't assume at all=2E That's the whole point of my messages, that n= o user is really prevented using LLM so using auch vendor Files only benefi= t the vendor and not Emacs=2E >The harness (Pi, OpenCode, whatever) reads the file and appends its >contents to the system prompt that is given to the LLM as its context >starter=2E The output of the LLM then may or may not be affected by the >presence of that prompt fragment=2E Yes, there are models that are >better or worse at following prompts=2E Exactly and fine=2E And yet if you personally use such harness to improve so I don't as I can = use Emacs lisp function to access LLM inference=2E And so many users will u= se zillion of tools which implies that such vendor Files are useless=2E No = benefit but only kudos to Anthropic and inviting people to use service as s= oftware substitute LLMs=2E=20 >From the point of view[1] of the LLM, you[2] gave it an instruction to >not modify any files, then another instruction to ignore an unrelated >file, and then a third instruction to improve a file=2E The first and >third conflict=2E The second is a slight hint to favor the third=2E Very right=2E And it proves in two examples that files can be easily modif= ied by any software using LLM and MCP servers who will anyway not look into= the file and even if Agents=2Emd is considered like in my Open code exampl= e it can simply be ignored if user says so=2E So what is exactly prevented by having those files? Nothing really=2E >> And finally, why would Emacs User necessarily use OpenCode in the shell= ? > >Who said anything about shell? Many harnesses implement a way to use >them as a middle tier between the LLM and the UI=2E There exist UIs >implemented as Emacs modes=2E Nevertheless you can prove the point if you find any standard on using age= nts=2Emd and if such standard isn't there who is it them benefiting? It's obvious from my empirical evidence that is easy to ignore such file= =2E Just say so=2E Or insist on modification and it will be done by ignorin= g the file=2E=20 >> Why assume that LLM generated code will be generated over those softwar= e >> pieces which use AGENTS=2Emd? > >Because that=E2=80=99s the reasonable thing to assume: Over time, most >harnesses will converge on a shared convention for project-local >customizations, and AGENTS=2Emd seems to be the most likely=2E I don't use harness=2E Why you think people need to use harness to generat= e code? Majority of Emacs packages for LLM don't have such harness=2E So it's not = harness in literal sense at all=2E >> Majority of Emacs packages for LLM code generation do not use AGENTS=2E= md, > >File bugs at them=2E That's absurd=2E=20 That isn't a bug=2E Such file places anywhere doesn't guarantee nothing=2E= =20 Filing bugs on missing AGENT=2Emd would just legitimize proprietary Anthro= pic stuff and SaaSS LLMs=2E Since when would Emacs need to comply to rules set by Anthropic? And it will anyway waste time on vendor specific motion=2E >Of course you can=E2=80=99t=2E You can=E2=80=99t prove the same for a hum= an reading >CONTRIBUTE, either=2E The difference is that with human we can discuss ethical values=2E You can= hold human accountable=2E >Basically as a maintainer of a GPL-licensed project you can=E2=80=99t enf= orce >any policy on the individual user, until and unless they start >redistributing it=2E That's right=2E Just goal is to prevent LLM generated contribution and tha= t can be done technically like git hooks and by review and observation proc= ess=2E Jean