learnaiwithrafa
ChatGPTWorkflows

The Three Codex Upgrades Actually Worth Your Time

I run both Claude Code and Codex. Codex just had a big month — here are the three changes worth the ten minutes it takes to set them up, with copy-paste prompts and the one myth to drop.

4 min read4 sources
  • #codex
  • #openai
  • #agents

I keep both Claude Code and Codex open, and I reach for whichever fits the job in front of me. Codex had a big month, and three changes earned a permanent spot in how I work. OpenAI says 5 million people now use Codex every week — and the fastest-growing slice, about 20% of them and growing more than 3x as fast as everyone else, aren't developers. That's the tell: Codex is becoming a place you hand off whole jobs, not just code.

1. Plugins: stop starting from a blank agent

OpenAI shipped 6 role-based plugins — Data Analytics, Creative Production, Sales, Product Design, and two finance ones — bundling 62 apps and 110 skills between them. A plugin pre-loads the agent with the tools a specific role needs. In the CLI you open the directory with /plugins; to call one directly you type @ plus its name. Under the hood a plugin can carry skills, apps, MCP servers, hooks, even scheduled-task templates — so it's less "an integration" and more "a coworker who already knows the tools."

Turn one on, then give it a real brief:

@Creative Production Here's the brief: I'm launching one product and want a
cohesive ad set. Pull a quick mood board, then take my product photo and produce
6 display variations sized for an Instagram feed post, a Story, and a web banner.
Logo bottom-left, clean and modern. Show me a review version before anything final.

2. Sites: ship an internal tool in one prompt

Drop the myth that you deploy with an @Sites command — there's no such thing. Sites lives in the ChatGPT app (chatgpt.com/sites), and publishing is a deliberate two-step: Save a version to review, then Deploy a version to a real production URL. It'll even back a simple app with a database (D1) and file storage (R2). A prompt I actually reuse:

Build a lead-capture page for my consultancy. Fields: name, work email, and
"what do you need help with". Store every submission somewhere I can view and
export, and show a short thank-you after submit. Save a version for me to review
first — don't deploy yet.

3. Scheduled tasks: work that runs without you

This is the one that changed my week. In the Codex app, Scheduled tasks let you set a recurring job in plain language or with an RFC 5545 RRULE like "every weekday at 8am". It runs in the background — in your local project or a Git worktree — and drops the result in your inbox. The CLI can manage these but won't run them interactively; the power-user path is /goal, where you hand off a big objective and let it loop until done. Set one up:

Every weekday at 7am: summarize what my three main competitors shipped or posted
in the last 24 hours and any public mentions of us. Give me the 3 highlights and
the one thing I should react to today. Drop it in my inbox.

Do this today

Turn on one plugin you'd genuinely use every week, then create one Scheduled task — the 7am competitor-and-mentions digest above is the highest-leverage first one. None of this is about a smarter model. It's leverage: less time typing in the box, more work happening while you're doing something else.

Sources