OpenAI Gave Us a Glimpse Into Their AI Coding Playbook
By Katie Parrott / Source CodeAbout this book
Summary
Was this newsletter forwarded to you?Sign upto get it in your inbox.Four OpenAI engineers built the Android version of theimage generation app Sorain 28 days. Naturally, they built it usingCodex, their AI coding agent.Partway through that sprint, one of those engineers,RJ Marsan, shattered his wrist in a bike race, leaving him unable to type. So he cobbled together a speech-to-text system and started talking to Codex instead of typing commands himself.That forced constraint—having to tell the computer what to do rather than execute the instructions via his keyboard—taught him something the rest of the team eventually adopted: Treating Codex like a new coworker you’re onboarding delivers stronger results than treating it like a tool you’re configuring (Codex doesn’t remember previous conversations automatically). “Every session is onboarding this new coworker [anew].”Marsan andAlexander Embiricos, who leads the Codex product team, joined us for Every’s first-ever Codex Camp to share this and other insights learned building Sora with Codex and how generally to think about working with AI.Here’s what we learned.Key takeawaysOnboard your AI like a new hire.Start with quick, interactive prompts. Build trust. Let it learn your preferences. Then delegate longer tasks.Don’t overload context.If you’d overwhelm a coworker with 6,000 facts about your codebase, you’ll overwhelm the AI too. Give it what it needs for the task at hand.Narrow beats broad.An agent with one focused job outperforms a generalist trying to catch everything.It doesn’t get easier—you go faster.AI tools shift the bottleneck; they don’t eliminate it. Architecture and code review become more important, not less.How to think about working with AIBuilding Codex changed how the OpenAI team thinks about AI agents...Become apaid subscriber to Everyto unlock this piece and learn about:Why OpenAI’s team created three specialized AI “police” roles instead of one catch-all code reviewerThe counterintuitive discovery about how much context actually helps—or hurts—an AI agentThe research step most developers skip—and why it causes AI-generated code to fail in hard-to-trace waysSubscribeClick hereto read the full postWant the full text of all articles in RSS?Become a subscriber, orlearn more.Book information
Genre
Business and Economics
Length
5 mins
Publish date
Dec 19, 2025
Language
English