Codex Tutorial for Beginners (Full Course)
TL;DR
Codex is pitched as a unified AI agent workspace, not just a chatbot — Riley Brown frames it as one interface for coding, browser research, file creation, automations, plugins, and even full computer control, then spends 1 hour 43 minutes showing that range with spreadsheets, docs, Swift apps, decks, and video.
The core workflow is project-based: every chat lives in a folder and every output becomes a real file — he creates a “Codex Desktop Research” project, has Codex research the new desktop app, generate an .xlsx spreadsheet, edit it in place, and then @-mention that file from a new chat inside the same project.
Plugins, skills, and automations are where Codex becomes practical for work — Riley connects Google Calendar and Gmail, asks for a weekly recap email, then turns that into a Friday 4:00 p.m. automation; later he builds a custom YouTube transcript skill using the Supadata API and schedules a monthly report.
The big lesson in part two is multitasking, not just prompting — Riley runs parallel chats to design an iOS app called Chorus, build the Swift app in Xcode, create a React landing page, draft an investor deck, generate a Remotion launch video, and wire up social posting automations, arguing this is the real skill as agents now work for 1-2 hours per task.
Codex is strong on orchestration, but Riley repeatedly reaches for Claude when design quality matters — he says Codex’s Figma integration is weak compared with Paper, calls some visual output “hideous,” and uses Claude Code inside the same project folder to polish the landing page, launch video scenes, and investor deck.
By the end, he has a rough but functioning business stack, not just a demo app — Chorus gets Swift screens, a Supabase-backed database, email/password auth, an app icon, a TestFlight build, a live Vercel waitlist site using Tally, a Canva-exportable investor deck, a Remotion promo video, and three automations including Typefully X drafts.
The Breakdown
Codex starts like ChatGPT, then immediately turns into a file-based agent
Riley opens by calling Codex OpenAI’s “super app” and says the big unlock is that it doesn’t just answer questions — it works inside a real folder on your machine. He creates a project folder, runs two research chats in parallel, and shows how Codex can turn web research into an .xlsx spreadsheet that lives in an outputs folder you can open in Finder.
The little UI details that make the workflow click
He pauses to explain the mental model: project location, permissions, model, effort level, and the side panel are the real basics. His default setup is “full access,” GPT-5.4, and “extra high” effort — in his words, just “let it cook” — and he likes how you can jump between chat view and a full-screen document preview, then comment directly on the artifact.
Search, plugins, and the blurry line between “skills” and “connectors”
Riley shows off Command-G search to recover old chats from removed projects, then gets into the plugin/skill confusion that every AI tool seems to name differently. His practical explanation is better than the taxonomy: if you want Codex to check Gmail or Google Calendar, you have to give it that capability first — and once connected, he has it list his events, email him a recap, and convert the whole thing into a weekly automation in one thread.
Building a custom YouTube transcript skill from an API key
One of the stickier moments is him solving an annoying recurring task: pulling YouTube transcripts at scale. He asks Codex to research transcript APIs, picks Supadata, pastes in an API key off-camera, invokes a “skill creator,” and ends up with a reusable “YouTube researcher” skill that summarizes Riley Brown’s last 10 videos, compares hooks, and outputs a document with thumbnails.
Part two: the real tutorial is how to juggle six agents at once
Riley explicitly says the point is not normal multitasking but serial focus: write one strong prompt, send it, then move to the next thread while the first one works for 5-10 minutes. He creates a new business folder for an app called Chorus and starts parallel chats for the project plan, iOS screen design, a Swift app scaffold in Xcode, a landing page, an investor deck, a launch video, and later social automations.
Chorus goes from prototype screens to a real Swift app with Supabase auth
He uses a custom mobile design skill based on Anthropic’s new design tool, gets a clickable prototype, then asks Codex to implement those screens in Swift. From there he iterates in the iOS simulator and on his physical iPhone, adding blur effects, saved states, platform listings, app icons, a Supabase-backed database, and finally email/password auth after abandoning Google sign-in as too much friction for the moment.
Where Codex shines, where it doesn’t, and why Claude keeps sneaking in
The most honest sections are the design ones: Riley says Codex’s Figma integration is mostly useful for turning Figma into code, not generating good design, and he strongly prefers Paper for AI-first layout generation. Later, when the landing page and Remotion scenes get clunky, he opens Claude Code from the terminal inside the same project folder, saying flatly that Claude is better at design-heavy tasks even though Codex is the better overall interface.
The final stack: app, site, deck, launch video, and posting automation
By the end, the rough startup stack is actually real. He deploys the React waitlist page to Vercel with a Tally embed, exports the investor deck to Canva for cleanup, gets Chorus into TestFlight, and creates a Typefully control skill so Codex can draft three X posts every morning — ending with the main takeaway that if you get stuck, ask the agent, and if a capability is missing, ask it to build the skill.