How to Use Codex Like an OpenAI PM | Abhi Muchhal, PM OpenAI (ex-Meta and Nubank)

Today’s episode

Six months ago, I told you Codex is the best way to use ChatGPT for PM work.

Most of you tried it. Some of you stuck with it and very few of you are running it the way the people who built it actually run it.

Today we get that inside look. Abhi Muchhal is an International Growth PM at OpenAI. Before that, Meta, Nubank, and a founder building on the OpenAI API. He is one of the people responsible for ChatGPT’s growth in India, Brazil, and Japan, markets that are now driving a meaningful share of OpenAI’s 900 million weekly active users.

He opened his actual setup on camera. The harness. The automations. The prompts that actually work. And the ones that failed before he figured it out.

----

Brought to you by:

Bolt.new - Ship AI-powered products 10x faster

Product Faculty - Get $550 off their #1 AI PM Certification with code AAKASH550C7

Customer.io - Send smarter messages using your product data

Ariso - Ship AI agents and features faster, with fewer regressions

Jira Product Discovery - Plan with purpose, ship with confidence

----

If you want access to my AI tool stack - Dovetail, Arize, Linear, Descript, Reforge Build, Relay.app, Magic Patterns, Speechify, Bolt.new and Mobbin - become an annual subscriber ($150), and grab Aakash’s bundle.

If you want access to my AI PM customizations - PM OS, Job Search OS, and Prompt Library - become a founding subscriber ($250).

----

Key Takeaways:

1. The harness is what separates Codex users from Codex runners - The connectors, the permissions model, and the skills layer are the three components that make Codex a system rather than a chat tool. Without all three, you are using an expensive autocomplete.

2. Generic prompts hit the wrong data - Abhi's team had separate B2C and B2B tables that both matched "tell me about weekly active users." The generic query returned the wrong answer every time. Specificity is the skill, name the exact dashboard and the exact metric, looks simple but saves a lot of time when you scale.

3. Three permission levels - Read tasks get full autonomy. Synthesis and drafts get full autonomy. Anything going to another human gets your eyes first. Treating permissions as binary, all control or all autonomy, breaks.

4. The person who cares most builds the skill - One OpenAI growth team built a skill that automates their entire experiment review process. It writes the hypothesis, monitors the run, and prepares the review doc.

5. Real automations run without you - Abhi runs three automations before he opens a single dashboard: a Slack triage, a 9:30AM self-refreshing growth dashboard pulling from 7-8 sources, and a weekly stakeholder update that writes its own first draft. He reviews, makes edits if needed, and sends.

6. Prototype before you document - Build the working prototype first, then write the 10-question companion FAQ. Showing engineers something that runs changes the conversation from whether to build to how to build it.

7. India is OpenAI's second largest market and under 10% of working adults are knowledge workers - The ChatGPT use case that drove US growth does not reach the same share of people in the markets driving the most new users. Building for the world means knowing how different the world actually is.

8. The WhatsApp computer use loop ran in 68 seconds - Point Codex at the WhatsApp desktop app. It reads what you missed, identifies action items, checks your calendar, and types the draft in the composer. One tap to send. Every PM building for international markets should run this workflow at least once.

9. Speaking evals is the key to breaking into a frontier lab - Name a capability you care about. Describe how you would measure it. Say how you would know if the model improved. You do not need 50 evals under your belt. You need to understand why they exist and what a good one measures.

10. Building something real is non-negotiable for frontier lab applications - Abhi had a live Chrome extension running on the OpenAI API at the time of his application.

----

Related content

Podcasts:

The Ultimate Guide to ChatGPT Codex

How PMs Ship 100K Lines of Code at OpenAI

Evals are the new PRD

Newsletters:

OpenAI’s Claude Code Killer

AI Agents Guide for PMs

How to Land a $300K+ AI PM Job

----

Where to find Abhi Muchhal:

LinkedIn: https://www.linkedin.com/in/abhimuchhal/

OpenAI:LinkedIn: https://www.linkedin.com/company/openai/W

here to find Aakash:

X: https://x.com/aakashgupta

LinkedIn: https://www.linkedin.com/in/aagupta/

Newsletter: https://www.news.aakashg.com

---

PS. Please subscribe on YouTube and follow on Apple & Spotify. It helps!

If you want to advertise, email productgrowthppp at gmail.

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.news.aakashg.com/subscribe