Digital resources in the Social Sciences and Humanities OpenEdition Our platforms OpenEdition Books OpenEdition Journals Hypotheses Calenda Libraries OpenEdition Freemium Follow us

AGENTS.md is an elevator notice

The instruction plate inside an elevator cabin, in French
The notice inside an elevator cabin, bolted to the wall where you enter. It says, in French: monitored around the clock; if stuck between floors, hold the alarm button and wait, someone will answer; and in red, ground floor has priority. Photo by the author, September 2026.

1,162 words, about 6 minutes.

I ended my last post by saying that if you work with a dragon you need a harness, and that yours need not be as elaborate as the Imperial Dragon Harness. We left a useful question unanswered: what is the smallest thing that counts as one? Smaller than you think, and it can already help to do the one job that matters most to a researcher: keeping invented references out of your manuscript.

A first prerequisite: keep project folders. Dumping all your work on your desktop may work for a while, but it quickly makes it nigh-impossible for your collaborators to work with you. An AI agent is a collaborator. So what follows assumes you group all the files related to a project into a project folder.

That cleared up, the minimal harness is one text file named AGENTS.md, placed in the project folder. That file is the sibling of README.md, but for an agentic instead of human readership. Here is the 101 tutorial:

The notice in the elevator

Look at the hero photo, the plate screwed to the wall of an elevator cabin.

It is short, because nobody reads a long notice. And it is written for a competent stranger. The notice does not teach you how an elevator works; it assumes you can press a button. It tells you what you could not have deduced from standing there: the cabin is monitored around the clock; if it stops between floors, hold the alarm button and someone will answer; and, in red, the instruction that outranks the others — the ground floor has priority.

AGENTS.md is that notice, posted where you enter the project. It is a plain text file. An agent such as Claude Code or Codex reads it automatically, for orientation, before it looks at your files and before you have typed a word. Your project folder is wherever the work lives. It need not be a code repository: it can be a manuscript folder, or a house-management folder. A one-page harness can be enough to describe how you work with documents in that folder.

The name AGENTS.md is a shared convention, so most AI agents should be able to use it. Including Claude Code since September 18th.

The .md ending says the file is written in Markdown: plain text, with # for headings and - for lists, as in the example below.

Example

Here is a thirteen-line harness example. It is organized in three blocks, in the order a visitor needs them: what this place is, what the rules are, how the work gets checked.

# AGENTS.md

## What this is
My thesis chapter on <subject>. Manuscript `chapter3.docx`, data in `data/`.

## Rules
- Cite only sources I have given you, and name the file and page each
  citation comes from. If a reference seems missing, tell me and stop.
- Treat `data/` as read-only. If a task seems to need a change there, show
  me the change and wait.
- Every number in the text comes from the analysis output: tell me which
  output line. If it disagrees with an earlier draft, say so instead of
  picking one.

## How to check your work
Show me what you changed before saving over anything.

You need not write all of it yourself. Draft the Rules block, since it is where you specify what you would otherwise have to repeat and repeat. Then ask the agent to draft the rest: it will visit the project folder once, summarize what it finds, and check your rules against it.

Such a notice buys what you most need at the start: the assistant stops guessing. Dropped into an unfamiliar directory with no notice, it would infer your conventions from what it can see, which is a polite way of saying it invents them. That is what any competent stranger does in a building with no signs; Borges wrote a whole library about it.

Notice what those rules share: each is a positive sentence, and each asks for a proof I can test, a file and page, a change shown before it is made, an output line. A notice should not forbid (see why below). It can ask to verify and to show. “Be rigorous” is a mood, and moods do no work. “Make no mistake” is a running gag; only verification works.

Limits

The common practice is to start with a small file, and add to it when you see where the agent needs steering. There are limits.

  • One is size. An AGENTS.md that runs over 150 lines is a sign that it is time to take the next step in harness writing.
  • Adding code snippets, detailed step-by-step procedures, including identifiers / passwords or keys in the file are short term conveniences one may regret later.
  • When rules pile up one incident at a time, eventually two of them will contradict each other. The model will then pick one, at random. You have built the Place That Sends You Mad, and your agent is queuing for permit A38.
  • The last limit is the one many tutorials skip. A notice does not stop anyone. People take the elevator during the fire alarm. What stops them is a locked door or an interlock in the mechanism.

An AGENTS.md rule is a notice in that exact sense. It is a polite request, if not a prayer. The model can overlook it, misread it, or decide that this case is the exception. These files are “context, not enforced configuration”, with “no guarantee of strict compliance” (Anthropic documentation).

The notice is only one of the several instruments available to steer the model: the three of my last post, plus one. Match the instrument to the cost of a violation.

Cost of a violation Where the rule belongs
A habit you would rather it kept AGENTS.md
A procedure you use often a skill, loaded on demand; it can call scripts
Any procedure that needs strict execution a small script
A corrupted file, a lost afternoon a hook: code that runs whatever the model decides

Writing PLEASE in capital letters is not a step up that ladder; it is the same instrument, shouting.

This afternoon

Open the folder of your current project. Create AGENTS.md. Write the three headings. Fill the Rules block with the two or three things you have already had to say twice this month. Claude Code and Codex will read it on their own. If you only use ChatGPT or Claude in the browser, paste the file at the top of each new conversation, or into the instructions of a Project: same notice, different door.

It is the simplest harness there is, a rope rather than the full imperial gear. It will still keep the dragon pointed your way while you learn to ride.

Download this post as a PDF → https://minh.haduong.com/files/HaDuong-2026-agents-md-elevator-notice.pdf

Minh Ha-Duong
Minh Ha-Duong
Directeur de Recherche, CIRED/CNRS
Minh Ha-Duong is a senior researcher at CNRS and CIRED, working on the economics of climate change and the energy… Read more

OpenEdition suggests that you cite this post as follows:
Minh Ha-Duong (September 17, 2026). AGENTS.md is an elevator notice. Research excellence à la Française. Retrieved October 6, 2026 from https://realf.hypotheses.org/200


2 comments

  1. Pro Tip: now that this article exists, I can refer this article to my favorite AI for it to fetch the advice and propose me a good agents.md that fits the above rules 😉 !

Leave a comment

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.