Author: Alexander Embiricos
Type: podcast
Published: 2026-01-12
Status: unread
Tags: source, ai-pm, claude-added

3 Advanced Codex Workflows: Git Worktrees, Plans.md, and Automated Code Reviews

By: Alexander Embiricos Host: Claire Vo Source: How I AI (ChatPRD / Lenny’s Podcast Network) Type: podcast

Summary

Four workflows from beginner to advanced. Workflow 1 (Zero-to-One): Use Codex in VS Code to understand unfamiliar codebases (ask “how do I run this?”) and make simple changes via natural language (e.g., “jump is way too big, lower please”). Codex identifies relevant code, forms plan, presents diff. Workflow 2 (Git Worktrees): Codex creates worktrees from terminal commands for parallel isolated development — two instances work simultaneously on different branches without merge conflicts. Workflow 3 (Plans.md): Meta-planning technique from OpenAI blog — create a Plans.md template that instructs the model how to structure implementation plans (self-contained milestones, edge cases, updating as it works). Feed task + Plans.md → detailed multi-step plan → review → execute. Used to build Sora Android app in 28 days. Plans can be 120+ lines. Bonus (Automated Code Review): Codex integrated into GitHub PRs — auto-scans changes, flags only high-confidence issues, engineers can reply “fix it” and Codex pushes a commit. Calibrated to protect scarce human attention. Used on almost every OpenAI repo.

Key Ideas Extracted

  • Engineer as architect/curator, not author: Role shifts from writing code line-by-line to reviewing, directing, and curating AI-generated work at a higher level of abstraction
  • Plans.md as meta-plan: Give the model instructions on HOW to plan (structure, milestones, edge cases) rather than just WHAT to build — creates more reliable, reviewable plans
  • Git worktrees + parallel agents = multiplied throughput: Multiple isolated environments let you prototype conflicting ideas simultaneously without merge conflicts
  • Protect scarce human attention in code review: Automated reviews should only flag high-confidence issues — false positives waste more human time than they save
  • In-thread fixing closes the loop: Reply “fix it” to a Codex review comment → new commit pushed — entire review-fix cycle happens in GitHub without switching contexts
  • Codex accessible to non-engineers: PM or designer can ask “how do I run this?” and make natural-language changes to unfamiliar codebases — lowers the barrier to contribution
  • AI handles “developer toil”: Complex git commands, boilerplate setup, and tooling configuration — Codex manages these so developers focus on decisions, not mechanics
  • 28-day shipping velocity: Small team built entire Sora Android app using Plans.md technique — structured planning + AI execution enables extreme speed

Notes

  • Published Jan 12, 2026 on ChatPRD blog. 9-min read.
  • Tools: Codex (VS Code extension, terminal, GitHub integration), Git worktrees
  • Codex included with ChatGPT Plus plan
  • Pro-tip: drag Codex icon to VS Code secondary sidebar (right side) for easier access
  • Plans.md template available on OpenAI blog
  • Sponsors: Brex, Graphite (AI code review platform)
  • Cross-reference: Lenny’s Newsletter post “Why humans are AI’s biggest bottleneck” (same guest)

Raw Content

Re-scraped from ChatPRD 2026-02-15. Full article content captured in Summary and Key Ideas above.


This site uses Just the Docs, a documentation theme for Jekyll.