Re: AGENTS.md and CLAUDE.md Are Concessions to Proprietary Services, Not Defenses Against Them

On Wed, Aug 5, 2026 at 2:33=E2=80=AFPM Dr. Arne Babenhauserheide
<[email protected]> wrote:

> One thing I don=E2=80=99t understand in this discussion: don=E2=80=99t th=
e tools check
> the CONTRIBUTE file?

Not automatically. If the user specifically tells it to, it probably
will. If the user has a personal prompt customization like =E2=80=9Cfor any
repository you work with, consult files with names in CAPS in its root
directory=E2=80=9D, it probably will, too. *But* users mostly don=E2=80=99t=
.

On the other hand, many tools *will* read AGENTS.md in the project
root and put its contents in the prompt.

> Could someone with existing tooling check whether adding a section about
> LLMs in CONTRIBUTE is adhered to by code generation? We have notes
> about git in it, so a note about other tools may be placed there, too.

LLM agents do not actively read every file in the project.

There is a system prompt that varies by harness. There are
user-specific additions to the system prompt; where they live is also
harness-specific. There is the project-specific addition to the
prompt, with many harnesses converging to AGENTS.md in the project
root. And then there are actual user prompts where the user states
their immediate task.

The LLM receives all of the above prompts, and then generates its
response. In typical agentic development setup, that response may
contain text for the user, and/or tool calls that the harness will
execute. (Some harnesses, for some tools, ask the user for permission
first.)

A user request typically results in a tool call to a shell command
like find =E2=80=A6 -e grep =E2=80=A6 with some terms derived from the user=
 prompt,
then read calls on files that are found. Then, depending on the model,
the prompts, and the contents of whichever files the model reads,
either the answer to the user=E2=80=99s question, or tool calls to modify s=
ome
files, or further exploration of the project.

Therefore: Putting project-specific LLM prompt customizations in
AGENTS.md has the greatest probability that they will be noticed.

> Also AGENTS.md is in markdown while CONTRIBUTE is in outline format.
>
> Do the tools adhere to files written in outline format?

The format does not matter much. What matters is words. Markdown is
just one convention that makes formatting lightweight; outline and Org
will work just as well. Compare with HTML, where each formatting
boundary is at least one word; SGML-based formats would waste tokens.

Putting outline-specific formatting in a file named *.md would be
somewhat annoying for humans who edit those, though.