4 steps, no coding
The short answer
Four steps, about ten minutes, and no terminal or IDE at any point. Codex now runs inside the ChatGPT desktop app, so a PM can drive a real coding agent from a normal window with a file list on the right. The one step people skip is step three, and it is the step that decides how good the output gets. Last updated August 02, 2026.
George Nurijanian · 6 min read · Product Leader, 8+ years
Nothing here needs a developer. Do these in order and you will have a working agent pointed at your own project folder.
Install the ChatGPT desktop app
Codex is not a separate download any more — it lives inside the ChatGPT desktop app. Install that and you have it. If you already run Claude Code, the app offers to import that setup on the way in, so your existing configuration comes with you.
chatgpt.com/download ↗Open your project folder
Projects, then open a folder. Two modes sit in the top left: ChatGPT for documents, decks, research, and strategy, and Codex for the agent itself. Your files list down the right-hand side, and every change the agent makes shows up in Review as a visual diff.
Projects → open a folderSet reasoning effort to max
This is the step people skip, and it is the one that matters most. Max effort is not the default. Set it once per project. The same model at max effort thinks for longer and checks its own work, and the difference in output quality is larger than anything you will get from rewriting your prompt.
Set it once per projectCreate AGENTS.md if you also use Claude
Codex reads AGENTS.md. Claude Code reads CLAUDE.md. Rather than maintaining two sets of rules that quietly drift apart, put three lines in AGENTS.md pointing at CLAUDE.md. Both agents then follow one source of truth, and you only ever edit one file.
Three lines, one source of truthThis is the part that changed. Every capability below used to sit behind a command line, which is exactly where most product managers stopped. Now it is a window.
Files panel on the right
Your whole project folder, visible and clickable. Nothing to memorise and no commands to type just to see what is there.
Visual diffs in Review
Every change the agent proposes is shown before it lands. Red and green, side by side, the way a pull request looks.
The agent in the same window
No second app to alt-tab into. The conversation, the files, and the changes all sit in one place.
Autonomous loops with /goal
Describe the outcome you want and let it run. It plans, works, checks itself, and keeps going without you steering every turn.
Project skills and MCP
Connect the tools you already live in — Jira, Notion, Linear — so the agent pulls real context instead of asking you to paste it.
Manual compaction
You decide when the context window resets rather than having it happen behind your back. That is the main lever you have on cost.
The instinct is to reach for the most expensive tier, on the theory that PM work deserves the good model. For this kind of work that instinct is usually wrong, and it is worth understanding why.
Subscription pricing is heavily subsidised compared to paying per token through the API. A flat monthly plan buys far more real work than the same money spent on metered API access, which means the cheap tier is not the compromise it looks like.
What actually limits output quality is the reasoning effort setting, not the price of the plan. An entry-tier subscription running at max effort will out-think a premium plan left on its defaults. That is the trade nobody puts on the pricing page.
Prices and plan limits move constantly. Check the current tiers before committing to one.
Plenty of PMs end up running Codex and Claude Code on the same project. The trap is maintaining two rule files that slowly disagree with each other, so that each agent behaves differently on the same task and nobody can work out why.
The fix takes thirty seconds. Codex reads AGENTS.md at the repo root. Point it at CLAUDE.md and stop there. Keep it tiny — the goal is behaviour, not ceremony.
Paste this into AGENTS.md
This project's agent and contributor guidance lives in CLAUDE.md. Read it before making changes. Do not add rules here, so there is a single source of truth.
This is also why AI PM OS works in either agent. The whole system is markdown — skills, frameworks, workflows — so there is no editor lock-in. Set the pointer once and the same 243 skills load whichever one you open that morning.
Setup is not the hard part. Getting one real result is. Pick something already on this week's list and run it through.
> /goal
Type it, describe an outcome, and let the loop run. This is the fastest way to understand what an autonomous agent actually feels like.
> Read this repo. Explain what it does.
The single best first prompt for a PM. You get a plain-English map of a system you are about to write specs against.
> Change something, then open Review
Ask for a small edit and look at the diff. Seeing changes as red and green lines is the moment this stops feeling like a chat window.
Setup, plans, reasoning effort, and running Codex alongside Claude Code.
No. There is no terminal and no IDE involved. You open a folder, type in plain English, and review changes as visual diffs. The skills that matter are the ones you already have — describing an outcome clearly and telling good work from bad.
For PM work, yes. Subscription pricing is heavily subsidised compared to paying per token through the API, so a monthly plan stretches much further than the sticker price suggests. The thing that limits output quality is the reasoning effort setting, not the plan tier.
How long the model thinks before answering, and how much it checks its own work. At max effort it will read more of your project, consider more alternatives, and catch its own mistakes. It is slower per task and dramatically better on anything that needs care.
Yes, and it is a reasonable way to work. Codex reads AGENTS.md and Claude Code reads CLAUDE.md, so point one at the other and both agents follow the same rules. AI PM OS ships as markdown skills, which means it loads in either one.
A browser chat forgets your project the moment you close the tab. A folder does not. Every document you produce becomes context for the next task, and the agent can read and edit those files directly instead of asking you to paste them in.
Anything where you cannot judge the output. It is very good at drafting, synthesising, and exploring a codebase, and it is confidently wrong often enough that unreviewed work will eventually embarrass you. Read the diff.
An empty folder and a good agent still leaves you writing every prompt from scratch. AI PM OS drops 243 PM skills, 150+ frameworks, and 11 guided workflows into that folder — as markdown, so Codex and Claude Code both read it.
Open AI PM OS →$99 one-time for individuals. Instant download via Polar.