Hey friends,
OpenAI has shipped a lot recently.
ChatGPT Work is here, GPT-5.6 has landed, and I’ve been giving Codex more serious work.
Before I show you the bigger jobs, I want to start with the small setup that makes those jobs easier.
It’s called AGENTS.md.
AGENTS.md tells Codex how you like to work before it begins.
Mine explains who I am, how I want technical work explained, what Codex can prepare, and what must come back to me for approval.
This is part one.
Later this week, I’ll share part two: 10 real jobs I give Codex as a non-technical operator.
For now, let me show you how to create your own AGENTS.md without writing it from scratch.
📌 What AGENTS.md Does
I created mine because I was tired of repeating the same working preferences across new tasks.
An AGENTS.md can tell Codex:
- Who You Are: Your work, experience and level of technical knowledge.
- How You Want It To Work: How it should explain things, use your files and report progress.
- When It Needs Approval: What it can prepare and what it must never publish, send or change without asking.
- What Finished Means: The checks it should complete before saying the job is done.
That gives Codex a better starting point.
You can still give instructions inside each task, but you don’t have to rebuild the working relationship every time.
✨ My AGENTS.md Has Two Jobs
Some instructions should follow me across Codex.
Others only make sense inside one workspace.
✅ 1. My Global Codex Instructions
I use a global AGENTS.md for the preferences that should apply everywhere.
Mine says I’m a non-technical operator.
Codex should explain technical work in plain English, preserve my existing work, and ask before anything public, destructive or difficult to undo.
This saves me from repeating the same boundaries whenever I open a different project.
Pro tip: Keep global instructions broad enough to help across all your work.
✅ 2. My Workspace Instructions
A newsletter and a website need different handling, so my main workspace folder has its own AGENTS.md.
It tells Codex to find the right project before working.
For a newsletter, it should follow my writing rules.
For a website, it should run the relevant checks before telling me the build works.
The workspace instructions add the details that would be too specific for my global setup.
Pro tip: Put project-specific rules close to the work they control. OpenAI explains how Codex layers AGENTS.md instructions here.
🛠 Ask Codex To Create Your First Version
You don’t need to know how to write an AGENTS.md.
Let Codex ask the questions and turn your normal answers into a first draft.
New To AGENTS.md?
Copy this into Codex:
I want to create an AGENTS.md tailored to the way I work.
Interview me one question at a time. Start by learning who I am and the work I do. Use my answers to understand how I want you to communicate, which actions require my approval, and what I expect before a task is finished.
After the interview, draft a concise global AGENTS.md for my review. Use plain language and include instructions that will help across my work in Codex.
Explain which details belong in a workspace AGENTS.md. Show me where you propose saving each approved file. Don’t create or replace anything until I approve the draft and location.
Already Have An AGENTS.md?
Ask Codex to review it against your real work:
Review my current AGENTS.md and recent work in this workspace. Find the instructions I keep repeating and the changes I keep asking for.
Show me which instructions should stay global and which belong in this workspace. Draft the changes for my approval.
Don’t replace my global or workspace AGENTS.md until I approve the wording and location.
That second prompt is close to what I used.
It was much easier than trying to remember every change I usually ask for.
🤖 Let Codex Read This Newsletter
This newsletter is written so your agent can use it too.
Copy the full issue, paste it into Codex, then add:
Use this newsletter as a guide to help me create my own AGENTS.md.
Interview me one question at a time. Draft the global instructions first, then explain which details belong in a workspace AGENTS.md.
Don’t save or replace any file until I approve the wording and proposed location.
You’ll still make the decisions.
Codex handles the interviewing and first draft.
📈 Why This Matters
A stronger model can do more work, but it still can’t guess how you like to work.
I now pay attention to the things I keep asking Codex to change.
One request may only matter for the current task.
If I make the same request again and again, I decide whether it belongs in AGENTS.md.
Over time, Codex starts closer to my preferences.
I spend less time repeating myself and more time reviewing useful work.
🫡 Final Thoughts
Start with the interview prompt.
You can finish today with a personal AGENTS.md draft that you understand and control.
Later this week, I’ll send part two with 10 real jobs I give Codex as a non-technical operator.
That episode will show what this setup looks like once Codex starts doing the work.
I’m also thinking about running a practical webinar on ChatGPT Work and Codex for non-technical users.
Reply webinar if you’d like me to let you know when it’s ready.
I’d love to know what Codex needed to learn about you.
To making Codex feel more like yours,
Nahid
