Learn 95% of Codex in 30 minutes
TL;DR
Codex is positioned as an AI-agent “super app,” not just a chatbot — Riley Brown frames it as OpenAI’s answer to tools like Claude Code, with local file storage, project folders, coding, document creation, image generation, browser control, and full computer use all in one interface.
The biggest practical difference is local file access — unlike ChatGPT or Claude storing uploads in the cloud, Codex reads and writes directly on your machine, which Riley demos by turning 53 receipt images into an Excel dashboard showing $25,982 total spend, category summaries, payment methods, and monthly trends.
Projects, memory, and reusable skills are the workflow backbone — Codex stores outputs in project folders, keeps a user-editable agents.md for manual preferences, maintains a separate auto-memory file, and can convert successful workflows like Gmail-based brand-deal analysis into reusable slash-command skills.
Plugins turn Codex into an operator across your tools — Riley shows it connecting to Gmail and Notion to scan two weeks of sponsorship emails, research brands, draft notes in markdown, and even write a new script in his own voice by studying previous long-form scripts in Notion.
Browser use and computer use push it beyond content generation into execution — with plugins, Codex opens Canva on Riley’s Mac, creates a five-slide presentation from generated sweater photos, then tests a browser-based app by clicking buttons, taking quizzes, and verifying navigation automatically.
Automations and Chronicle hint at a more ambient agent future — Riley sets a recurring Friday 9:00 a.m. automation to refresh his brand-deal spreadsheet, and shows Chronicle’s “always recording your screen” research preview using recent screenshots to suggest missing slides without him manually uploading context.
The Breakdown
Codex as the “easy interface” for AI agents
Riley opens by pitching Codex as OpenAI’s agent tool for both coding and knowledge work — basically Claude Code, but with a friendlier interface. His main framing is that this is not just another chat window: it can build landing pages, spreadsheets, presentations, apps, games, and even control your computer.
Full file access: the receipt demo that makes it click
The first big capability is local file access, and Riley contrasts it directly with ChatGPT and Claude storing files in the cloud. He drops 53 receipt images into a Downloads folder, asks Codex to OCR and analyze them, and seven minutes later gets a local Excel workbook with a dashboard showing $25,982 in spend, category summaries, payment methods, and monthly trends. The point lands because you can literally click “open in folder” and see the file sitting on his machine.
Projects keep the chaos manageable
Once local files are the default, Riley says the best way to stay organized is to work inside projects tied to folders on your computer. He creates a “seven capabilities” folder, shows how outputs only appear once the task is finished, and then demonstrates a neat workflow: @-mention a document in that folder and ask Codex to turn it into a landing page, with the generated code living in the same local project.
Memory: one file you should edit, one you probably shouldn’t
Riley splits memory into two types: manual memory in agents.md and auto memory in Codex’s hidden memories folder. He shows Codex remembering his preferred landing-page style — a text-first layout with a left sidebar and table of contents — by updating agents.md, then warns that the auto-memory file is more of an observation log you should inspect occasionally but not manually edit.
Plugins make Codex useful in the real world
Next he moves to plugins, describing them as installable connections to external apps and workflows. The Gmail example is especially concrete: Codex scans the last two weeks of his email, identifies sponsorship offers, researches the companies, and compiles a markdown table with notes; then a Notion plugin lets it read his old scripts and generate a new one in his voice, opening with, “Do you ever feel like AI coding tools are getting more powerful every week, but also somehow more confusing?”
Skills: from one good run to a reusable SOP
Riley’s core advice on skills is simple: don’t start with abstract instructions if you don’t have to. Have Codex do the task, iterate until the output is genuinely good, then say, “I’m happy with this output. Turn it into a skill,” which he does for a color-coded brand-deal spreadsheet workflow. He also shows an Excalidraw skill producing a polished visual outline, then immediately updates the skill after being pleasantly surprised by the formatting — every run is a chance to refine the SOP.
Images, browser use, and full computer control
Then the demo gets more agentic. Riley uses built-in GPT-image-2 to generate five product photos of models wearing his sweater, stores them in a local “content” project, and then invokes computer use to open Canva on his Mac and build a five-slide presentation with one image per slide. After that, browser use takes over inside Codex to test an app by clicking buttons, scrolling, completing quizzes, and checking whether navigation and side panels actually work.
Automations and Chronicle: recurring work and ambient context
The last formal capability is automations: once a task and skill work, you can just tell Codex to run it every Friday at 9 a.m., which Riley does for his brand-deal sheet. As a bonus, he shows Chronicle, a research-preview feature that continuously captures screenshots of your screen; he calls it a little invasive, but it does let Codex infer what’s in his presentation and suggest extra slides like a “Codex super app map” or a “why Codex over other tools” slide without any manual upload.