“A full software engineering teammate”: OpenAI product lead on getting the most out of Codex | Alexander Embiricos

Alex demonstrates installing Codex from the VS Code marketplace and shows basic workflows for non-technical users. He illustrates using Codex to answer questions about unfamiliar codebases, make simple changes like adjusting jump height in a game, and implementing missing features through natural language prompts.

• •Codex is included in ChatGPT Plus, Pro, Business, Team, and EDU plans with generous limits
• •Start by asking Codex questions about the codebase - useful for PMs and engineers alike
• •Use natural language prompts for simple changes without needing to understand the code
• •Codex excels at thorough, complex tasks rather than speed - it's the most diligent coding agent

Alex explains how to use git worktrees to run multiple Codex instances in parallel without conflicts. He demonstrates creating separate worktrees for different features and shows the command-line shortcut for launching Codex with prompts inline, enabling efficient parallel prototyping.

• •Use 'codex --' followed by your prompt to launch Codex with a command in one line
• •Git worktrees allow multiple copies of a codebase tracked by one Git instance
• •Ask Codex to create worktrees for you - no need to memorize Git commands
• •Parallel tasks work best when changes are unlikely to conflict; otherwise use serial execution

Alex reveals how OpenAI built the #1 App Store app (Sora for Android) in 28 days with 4 engineers using Codex. He emphasizes that success came from careful architectural planning rather than one-shot prompting, introducing the concept of using plans.md for complex projects.

• •Four engineers built Sora Android app in 28 days, immediately becoming #1 in App Store
• •With coding agents, development doesn't get easier - you just move way faster
• •Success requires thoughtful architecture planning upfront, not just vibing code
• •Use plans.md template from OpenAI blog for structured, milestone-based implementation plans
• •Plans should be self-contained with clear milestones that the agent updates as it progresses

Alex demonstrates creating detailed implementation plans using the plans.md framework for building a Python SDK. He shows how to iterate on plans within the same chat to maintain context, and explains when to use planning versus parallel exploration approaches.

• •Copy the plans.md template from OpenAI's blog and reference it when asking Codex to plan
• •Codex produces thorough 100+ line plans with naming conventions, milestones, and implementation details
• •Keep plan iterations in the same chat so the model has full context when implementing
• •For quick tasks, skip planning and run 4 parallel attempts instead using 'best of N' feature
• •Planning is essential for hard tasks; parallel exploration works for rapid prototyping

Alex showcases Codex's automated code review capabilities in GitHub, which only surfaces high-confidence issues to protect human attention. He demonstrates the closed-loop workflow where developers can ask Codex to fix issues it identifies, and discusses why some automation attempts failed.

• •Use /review command in Codex to get thorough code critiques before submitting PRs
• •GitHub integration provides automated code review on all PRs without manual prompting
• •Codex only comments when highly confident to avoid noise - silent approval means no issues found
• •Developers can reply to Codex comments asking it to fix issues, creating a closed loop
• •Automated PR revision based on human feedback had low hit rates and was discontinued

Alex shares that nearly all technical staff at OpenAI now uses Codex constantly, up from 50% earlier in the year. He discusses the 70% productivity increase measured by PR volume and explains the product philosophy of reducing friction by eliminating the need to type prompts.

• •OpenAI went from 50% to nearly 100% technical staff adoption of Codex in months
• •Early adopters showed 70% higher PR volume compared to non-users
• •Codex code review is enabled on virtually every repo at OpenAI
• •Product focus is on useful automation that requires zero user effort (like auto code review)
• •The biggest limiting factor for AGI benefits is whether people want to type the prompt

Alex explains why the harness (interface) matters as much as the model, discussing how Codex's open-source approach lets users see optimizations for each new model. He reveals features like compaction and parallel tool calling that maximize model capabilities.

• •OpenAI ships new models every two weeks - harnesses must evolve to maximize each model's strengths
• •Codex CLI is open source so developers can see and copy best practices for using OpenAI models
• •New capabilities like compaction allow models to start fresh conversations with optimized context
• •Many users prefer IDE integration over terminal-only workflows to see code changes in real-time
• •Product experience unlocks adoption - building the right UX matters as much as model quality

Alex shares his favorite Atlas (ChatGPT) features, emphasizing how memory creates personalized responses and how Side Chat enables contextual assistance on any webpage. He advocates for being polite to AI to protect human empathy and communication habits.

• •Tell ChatGPT about decisions you make based on its advice to improve future recommendations
• •Memory enables highly personalized responses (e.g., dinner recommendations based on location and preferences)
• •Side Chat allows asking questions about any webpage without leaving the page
• •Being polite to AI is important for maintaining your own humanity and setting examples for others
• •Use Cmd+T to open new tabs in Atlas for quick access to AI assistance

Alex shares advanced prompting techniques, emphasizing the importance of providing context and appropriate levels of ambiguity. He reveals that Codex stores session logs locally, enabling advanced users to have the agent read previous sessions when starting fresh conversations.

• •Always provide context and explain the 'why' behind requests, not just the 'what'
• •Product managers make great prompters because they're used to working with ambiguity
• •Don't create false precision - let the agent determine implementation details when you don't care
• •If stuck, start a new chat and ask Codex to read previous session logs from ~/.codex/sessions
• •Codex is hiring PMs and engineers - check the jobs page and reach out on Twitter @embirico