A map of the 5-hour Codex course: 18 core concepts, an AI operating system, skills, and a 15-task Codex vs Claude Code bake-off

The 18 core concepts (first 40 minutes)

If you watch one part, watch this. He sorts Codex into four groups:

• Foundations: 1 projects · 2 AGENTS.md · 3 the agent loop · 4 /goal.
• Environments: 5 local vs cloud · 6 worktrees · 7 Codex itself (how the desktop app, IDE extension and CLI relate).
• Control and customisation: 8 models · 9 effort · 10 permissions · 11 skills ("basically reusable workflows" — his favourite) · 12 agents · 13 plugins.
• Tools and scale: 14 the browser · 15 sites · 16 subagents · 17 scheduled tasks · 18 voice mode.

Install is simple: download the Codex desktop app for your OS and sign in with your ChatGPT account. You need at least a paid ChatGPT plan; he runs the $200 plan and shows the usage meter, which is the thing to watch.

The AI operating system (from ~41 minutes)

His central idea: a project folder that acts as a wiki the agent consults — sections for things like company structure, meeting notes, video knowledge and a voice/style reference in markdown — so every task starts with the right context instead of a fresh explanation. The AGENTS.md file tells Codex where each kind of knowledge lives.

Skills (from ~1h01m)

His method for building skills that behave the same way every time, then specific examples of skills tuned to one person's voice and brand. This covers the same ground as his six-step skills method, which we summarised separately — start there if skills are all you need.

Building things (1h41m – 3h25m)

• Images and video generation from inside Codex (~1h41m).
• Website design with GPT-6 Astra (~1h53m).
• Browser use — he rates Codex the best at it right now (~2h02m).
• Video editing, including subtitle overlays, with no editing experience (~2h19m).
• Connecting services such as GitHub and ClickUp (~3h05m).
• Voice mode for talking tasks through with the agent (~3h25m).

Codex (GPT-6 Astra) vs Claude Code (Fable 5.1): 15 tasks

Astra came out $186 cheaper but an hour and 43 minutes slower across the set. He adds a practical point about subscriptions: on his plan he isn't hitting a five-hour usage window with Codex the way he does with Claude, so he gets more work through in a week. Compare this with his later Opus 5.5 vs Astra test, where Opus 5.5 won 8–4 — the newer Claude model changes the picture.

Selling what you build (from ~4h20m)

The last hour is about turning this into client work: agreeing one KPI before accepting any payment, value-based pricing, and handling a client who balks at the number. He's candid that he priced many of his first hundred systems wrong.

Key Takeaways

• The 18-concept opening is the best 40-minute Codex primer we know of; use the chapter links to jump.
• Put your standing context in a wiki-style project and point AGENTS.md at it, so tasks start informed.
• Treat skills as reusable workflows and tune them to one voice and brand.
• On his 15 tasks, Astra won 10–5 at about two-thirds of Fable's cost, but ran longer — and Opus 5.5 has since changed the Claude side of that comparison.
• If you sell automations, fix the KPI and price on value before you build.