The power user’s guide to Codex | Alexander Embiricos (product lead)
Alexander Embiricos (guest), Claire Vo (host)
In this episode of How I AI, featuring Alexander Embiricos and Claire Vo, The power user’s guide to Codex | Alexander Embiricos (product lead) explores power user playbook for Codex: planning, worktrees, and reviews Codex is presented as a thorough coding agent that shines on complex tasks, used both for answering questions about unfamiliar codebases and for implementing changes via natural-language prompts.
Power user playbook for Codex: planning, worktrees, and reviews
Codex is presented as a thorough coding agent that shines on complex tasks, used both for answering questions about unfamiliar codebases and for implementing changes via natural-language prompts.
The episode demos a zero-to-one setup in VS Code and the terminal, then shows how to run multiple tasks in parallel safely using Git worktrees.
For larger efforts, Embiricos advocates structured planning (via a shared Plans.md template) and iterating on the plan in the same chat before executing implementation.
They highlight “harness” differentiation (product + workflow around the model), especially automated GitHub code review that catches issues with high confidence and can even fix them on request, plus broader integrations (web/cloud, Slack, Linear) and personalization via Atlas/Sidechat.
Key Takeaways
Treat Codex as a teammate for both questions and code changes.
Embiricos emphasizes that a high-frequency use case is simply asking questions—how to run a repo, what a feature flag does, or whether something shipped—reducing friction and avoiding unnecessary pings to engineers.
Get the full analysis with uListen AI
Parallelize work, but manage conflicts intentionally.
Small independent tasks can run in parallel chats; when changes might collide, switch to serial execution or isolate work using Git worktrees so each change remains reviewable and non-conflicting.
Get the full analysis with uListen AI
Use Git worktrees as the “safe parallelism” primitive.
Worktrees let one repo track multiple working directories (e. ...
Get the full analysis with uListen AI
For hard problems, planning beats one-shot prompting.
The Sora Android app story underscores that agents don’t remove architecture work; teams moved fast by defining the shape of the system, then using detailed plans as the contract Codex executes against.
Get the full analysis with uListen AI
Adopt a reusable planning spec (Plans.md) to standardize outcomes.
Copying a “how to plan” template into Plans. ...
Get the full analysis with uListen AI
Iterate on the plan in the same chat before implementing.
Keeping plan edits in one thread preserves context; once the plan is correct, you can hand Codex the plan file to implement with higher confidence and fewer reversals.
Get the full analysis with uListen AI
Automated code review is a force multiplier—if tuned for low noise.
Codex’s GitHub review automation posts only high-confidence issues to protect attention; users can reply “fix it,” creating a tight loop where the agent identifies and remedies problems without manual prompting.
Get the full analysis with uListen AI
Notable Quotes
“People love how thorough and diligent Codex is. It's not the fastest tool out there, but it is the most thorough and best at hard, complex tasks.”
— Alexander Embiricos
“We used Codex to build the Sora app for Android in twenty-eight days, and it immediately became the number one app in the App Store.”
— Alexander Embiricos
“With coding agents, it doesn't get easier, but you just move way faster.”
— Alexander Embiricos
“If you have a chat where Codex is producing these plans and you want to change something, it's actually really nice for the model if you just use the same chat to ask for changes to the plan.”
— Alexander Embiricos
“The bottlenecks are kind of like thinking about what code to write and then making sure that code is good, reviewing it and landing it.”
— Alexander Embiricos
Questions Answered in This Episode
In the Sora Android example, what did the team define manually (architecture/spec) versus what Codex generated, and what were the failure modes of the initial “one prompt” attempt?
Codex is presented as a thorough coding agent that shines on complex tasks, used both for answering questions about unfamiliar codebases and for implementing changes via natural-language prompts.
Get the full analysis with uListen AI
What are the concrete signals you use to decide “this needs Plans.md” versus “Best-of-N exploration is faster,” and how do you prevent wasted parallel attempts?
The episode demos a zero-to-one setup in VS Code and the terminal, then shows how to run multiple tasks in parallel safely using Git worktrees.
Get the full analysis with uListen AI
Can you show a real (non-contrived) worktree workflow: how you name branches/worktrees, sync changes back, and handle shared refactors across parallel efforts?
For larger efforts, Embiricos advocates structured planning (via a shared Plans. ...
Get the full analysis with uListen AI
How is Codex’s automated GitHub review tuned (confidence thresholds, categories of issues), and what types of bugs does it intentionally avoid commenting on to reduce noise?
They highlight “harness” differentiation (product + workflow around the model), especially automated GitHub code review that catches issues with high confidence and can even fix them on request, plus broader integrations (web/cloud, Slack, Linear) and personalization via Atlas/Sidechat.
Get the full analysis with uListen AI
You mentioned an auto-revise-on-feedback feature that wasn’t popular—what specifically made it miss (context gaps, nits vs design feedback), and what would need to change to revisit it?
Get the full analysis with uListen AI
Transcript Preview
people love how thorough and diligent Codex is. It's not the fastest tool out there, but it is the most thorough and best at hard, complex tasks.
If you're a software engineer or somebody who's even just new to using some of these AI tools, where would you get started with Codex?
We're building it into a full software engineering teammate. One of the things that Codex is great at is simply answering questions. If you have a chat where Codex is producing these plans and you want to change something, it's actually really nice for the model if you just use the same chat to ask for changes to the plan, and that way, it has all this context in its head when it's ready to get going.
This is a great starter flow that shows how flexible this platform is and how it can meet a bunch of people at a variety of levels of tasks. How is OpenAI using this for bigger features and bigger products?
We used Codex to build the Sora app for Android in twenty-eight days, and it immediately became the number one app in the App Store. [upbeat music]
Welcome back to How I AI. I'm Claire Vo, product leader and AI obsessive, here on a mission to help you build better with these new tools. Today, we have Alexander Embiricos, product lead for Codex from OpenAI, and he's gonna show us how you get the most out of Codex, whether you're a non-technical user trying to make changes to an existing code base or want the power tips and tricks for getting the most out of it in the terminal. Let's get to it. This episode is brought to you by Brex. If you're listening to this show, you already know AI is changing how we work in real, practical ways. Brex is bringing that same power to finance. Brex is the intelligent finance platform built for founders. With autonomous agents running in the background, your finance stack basically runs itself. Cards are issues, expenses are filed, and fraud is stopped in real time without you having to think about it. Add Brex's banking solution with a high-yield treasury account, and you've got a system that helps you spend smarter, move faster, and scale with confidence. One in three startups in the US already runs on Brex. You can, too, at brex.com/howiai. Alex, thanks for joining How I AI. I'm excited about today's episode because we actually haven't seen a deep dive into Codex yet, and we are gonna get the expert take on how to get the most out of this tool. And I love that we're just gonna dive in and do a zero to one "Hello, World!" with Codex. So if you're a software engineer or somebody who's even just new to using some of these AI tools, where would you get started with Codex?
Codex is a coding agent. We're building it into a full software engineering teammate. But to get started, let's just talk about where most people use it, which is in their IDE. Uh, I happen to use VS Code, so I'll show you Codex in VS Code. You can also use it, uh, the Codex extension in any VS Code fork, like Cursor, et cetera. So let's say that I just installed Codex from the VS Code extension marketplace. Do you want me to show that, by the way, or-
Install uListen to search the full transcript and get AI-powered insights
Get Full TranscriptGet more from every podcast
AI summaries, searchable transcripts, and fact-checking. Free forever.
Add to Chrome