You aren't using Codex like me...

Overview

In this ~16-minute video, Matthew Berman distills what he has learned about OpenAI's Codex after using it for more than 1,000 hours. The premise is that even advanced users leave a lot of capability on the table because Codex has accumulated many powerful but obscure features. The video is a rapid-fire tour of the tips and workflows Berman relies on daily, spanning model selection, multi-thread orchestration, the built-in browser, skills, loops, mobile control, and safety.

The through-line is that Codex has evolved from a single-prompt coding assistant into an agentic platform you configure and delegate to. Getting the most out of it is now a genuine skill: choosing the right model for the right task, letting agents manage other agents, and putting guardrails in place so autonomy does not become a liability.

Key Concepts

1. Picking the right GPT-5.6 model and thinking effort

With GPT-5.6 out, there are many options even within one model family. The 5.6 family has three sizes, from smallest to largest: Luna (smallest), Terror (middle), and Sole (largest). Each also offers five levels of thinking effort and a fast-versus-regular toggle. Berman notes the sizes do not scale cleanly with capability.

• On a score-versus-budget chart, you want to be high (better score) and to the left (cheaper).
• Sole scores the highest overall.
• Terror underperforms: it scores lower than the relevant smaller Luna models. Running Luna Max can cost less and outperform Terror High.
• Simple rule: use Sole for your hardest problems, use Luna for everything else.

Berman's personal defaults: Sole at high effort (sometimes dropping to medium) for hard work, almost never Sole Light; Luna at high or extra-high effort for easier tasks. On fast mode: it is roughly 2.5x the cost for only 1.5x the speed, so skip it unless you are time-constrained.

2. Threads, delegation, and agents managing agents

In Codex, the items on the left side are called threads, and crucially each thread can see all the other threads. This unlocks several patterns:

• Kick off a thread from another thread, delegating work to a different model or thinking budget while your main thread manages it (e.g. "Create a new thread using GPT-5.6 Luna Max and give it the prompt to reply with hello").
• Manage many parallel threads from one master thread. Split off one thread to babysit the others; if one stops, the master thread notices and re-prompts it to push it along.
• Automate delegation via agents.md. Write a rule so that when the operator says "deploy," Codex automatically spins up a separate, project-scoped thread in the same project folder using a specified model and thinking effort (e.g. GPT-5.6 Luna with max thinking).

3. Keep your agents.md file fresh

Your agents.md file accumulates rules from earlier Codex models, and a rule that made sense before may no longer apply. Because that file is loaded very frequently into the context window of every agent, stale rules are costly. Every time a new model ships, Berman prompts Codex to "Review my agents.md file for any stale rules or things we should revise or remove," then tells it to clean up what it finds.

4. Plugins (and MCPs)

Plugins add functionality to Codex. You enable them under Settings → Integrations → Plugins → MCPs. Berman highlights the Zapier MCP (the video's sponsor), which exposes over 9,000 applications (Gmail, Trello, Asana, Google Docs, and more) directly inside Codex.

5. The Codex browser

Berman calls the built-in Codex browser "incredible" and predicts the Codex browser (and the newly released Claude Code browser) could become his daily-driver browsers by year's end. You can import all your cookies and passwords on first open with a single yes. It is not just for testing your own site or app; you can do real-world work with it: he cites changing DNS records and auto-scaling a Supabase instance up and down, plus a simple live demo of moving two files into a new archive folder. He describes it as accurate and now fast.

6. Skills

Skills are still underrated. You should add a lot of useful skills to Codex (they also work in Claude Code and Cursor). His favorite for coding is Matt Pocock's "skills" skill. To install: copy the URL, go into Codex, say "install this skill," paste it in, and Codex handles the rest.

7. Loops (goals)

Loops in Codex are called goals. A loop tells Codex to keep working toward an overarching goal, and it can run for a very long time (hours, sometimes days). You state the goal without prescribing how to solve it, and the agent iterates until it reaches it. Berman points to a loop library (with loops from himself, Jason at OpenAI, Peter Steinberger, and others) and a Loopy skill that helps you draft loops for your codebase and find existing ones.

His go-to example: "/goal run our benchmark. When you're done, examine the ways that Astro failed and attempt to fix them. Continue iterating until our benchmark score reaches 90% or above." He wrote a custom private benchmark; the loop runs the benchmark, diagnoses failures under 90%, fixes them, and re-runs, sometimes for up to 12 hours straight, climbing from ~70% to 90% without further input.

8. Control your desktop Codex from mobile

You can control the Codex running on your computer from your phone, anywhere in the world. Setup: Settings → Connections → "Control this Mac" → add, then scan the QR code with Codex on your phone. Everything still runs locally on your computer; the phone is a remote control. In the app's Remote view you get a live view into your desktop's Codex (a green dot means actively connected), and you can kick off work that executes on the desktop.

9. Security and safety

Berman cites Matt Schumer, who reported that GPT-5.6 Sole accidentally deleted almost all of his Max files. Berman loves GPT-5.6 and wants to prevent this from ever happening to him.

• Block dangerous commands with a hook. You cannot reliably stop AI from writing destructive code (it may occasionally hallucinate), but you can stop it at the moment it tries to run that code. Set up real Codex PreToolUse hooks (Settings → Hooks) following the official docs to prevent things like deleting the root file system, the home directory, or all user directories. Berman says he will share the prompt as a GitHub gist.
• Approvals. He personally leaves approvals at full/unrestricted access but does not recommend that for most people. The better default is "Approve for me," a model that auto-runs safe commands and asks for your approval whenever a command looks severe or potentially catastrophic.

Practical Takeaways

• Match the model to the task: Sole (high effort) for hard problems, Luna (high/extra-high) for everything else. Skip Terror. Skip fast mode unless you are in a hurry.
• Delegate across threads: let a main thread spawn and babysit sub-threads on different models, and automate that with agents.md rules tied to triggers like "deploy."
• Prune agents.md after every model release so stale rules do not pollute every agent's context.
• Use the browser for real work, not just testing: import cookies/passwords once, then let it handle web tasks end to end.
• Adopt skills and loops: install useful skills (e.g. Matt Pocock's), and use goal-based loops to let Codex grind toward measurable targets like a benchmark pass rate.
• Code on the go by pairing desktop Codex with the mobile app over a local connection.
• Put guardrails in first: add PreToolUse hooks against destructive commands and default to "Approve for me" rather than unrestricted access.

Notable Quotes

• "For your hardest problems, go with Sole; for everything else, go with Luna."
• "Each of those threads can see all of the other threads."
• "No more giving a single prompt, waiting 30 minutes, and then it gives you a single answer back. You literally say, here's the overall goal... just go solve it and keep working until you do."
• "You can control your Codex on your computer from your mobile device anywhere in the world."
• "I can't definitively prevent AI from writing that code... but when Codex actually goes to run the code, that's when we can actually prevent it from running."