skills/plain-ai-explainer/references/approved-agents-md-thread.md

# Approved Fifth G Explainer Reference

This is the exact approved thread that established the `fifth-g-explainer` style

Do not rewrite, normalize, or silently improve this reference

### Hook (207 chars)

ChatGPT Codex can do your work

But it can also make the wrong call faster when it knows nothing about how you work

There is a fix and it's called `AGENTS.md`

It's also much less technical than it sounds 🧵

### 2/9 (365 chars)

An `AGENTS.md` is a set of written instructions Codex reads before it starts working

If you've heard of `CLAUDE.md`, this works the same way in Codex

You can use it to explain who you are, what kind of work you do, how you want Codex to speak to you, and which actions need your approval

You write those things once instead of explaining them again in every chat

### 3/9 (314 chars)

Mine says I'm a non-technical operator, so technical work needs to be explained in plain English

It also says Codex can research, check files, and prepare drafts on its own

But it has to ask me before it publishes, sends, or changes anything public

That's the kind of thing an agent can't guess from your prompt

### 4/9 (282 chars)

You can also tell Codex what “done” means

If it's writing a Thread, I want the draft saved in the right place and checked before I review it

If it's working on a website, I want the site to build, the links checked, and the page tested on mobile

Same agent

Different finish line

### 5/9 (268 chars)

The useful bit is that you can have more than one `AGENTS.md`

I have one main file for rules that should follow me across Codex

Things like how I communicate and when I want to approve an action

Then a workspace can have its own file for the work inside that folder

### 6/9 (293 chars)

So my main `AGENTS.md` can say, “Explain technical work in plain English”

The one in my work folder can say, “Find the project this task belongs to before you touch anything”

And a website project can add its own checks below that

Codex reads the instructions that match the work it's doing

### 7/9 (319 chars)

You don't need to sit down and write the perfect file from memory either

If you're starting fresh, Codex can interview you about your work and draft the first version

If you already have one, it can look through your recent corrections and show you which rules are missing

You still review it before anything changes

### 8/9 (248 chars)

This is the setup I think a lot of people will skip while they rush to test new agents

The model still matters

But it can't know how you work, what it can do without asking, or what a finished job looks like to you

That part has to come from you

### 9/9 (270 chars)

I ended up rebuilding my own `AGENTS.md` around those questions this week

I wrote down the setup I landed on, how I split the instructions, and the prompt that helped me do it

If you've been meaning to make your own, that's what this week's newsletter is about

[LINK]