The Best AI Coding Guides Don't Know Who They're For

Derrick Choi's Codex best practices guide opens with: "If you're new to Codex or coding agents in general, this guide will help you get better results faster."

By step three, you're managing AGENTS.md files across three directory levels. By step six, you're configuring OAuth-authenticated MCP servers. By step nine, you're running multi-agent threads against git worktrees.

That's not a beginner guide. It's product documentation wearing one.

Here's what it actually gets right. The AGENTS.md concept — a persistent file that encodes your standards, conventions, and definition of "done" so Codex doesn't need re-briefing every session — is a transferable idea. Most people treat prompts as disposable. The guide treats them as something worth investing in, capturing, and improving over time. That mental model is more useful than most of what's out there right now.

The rest is a feature tour. Useful if you're already in the tool, less so if you're deciding whether to invest in it.

What's conspicuously absent: any honest account of where coding agents still fall down. No mention of tasks where poor test coverage makes self-validation unreliable. Nothing about the classes of work — architectural decisions, anything with subtle business logic — where the cost of a plausible-looking wrong answer is high. The "start conservative with permissions" advice appears as a sub-tip buried in step four, well after the reader has been walked through automation and multi-agent parallelism.

Worth clicking through for the AGENTS.md section and the planning-before-coding framing. Skim the rest.

One more thing worth knowing: the author works at OpenAI. Not disqualifying, but it does explain why every principle eventually resolves into a specific Codex feature rather than something you could apply anywhere else.