Greg Isenberg opened his laptop, typed 6 prompts, and walked away with an internal app that updates itself.
No Replit. No Lovable. No dev team.
Just Codex Sites, a plugin inside OpenAI's Codex that most founders haven't even opened yet.
I watched the full walkthrough so you don't have to squint through 25 minutes of screen recording. Here's the entire playbook: what he built, the exact prompts he used, and the 5 concepts (memory, safe actions, skills, save gates, and proving the loop) that turn a toy demo into software that works for you 24/7.
Watch the source video here, then steal the process below.
Why Greg almost dismissed Codex Sites (and why he didn't)
His first reaction was the same one you probably had: "Is this just a worse version of Replit or Lovable that's stuck inside Codex?"
Fair question.
Replit and Lovable give you the whole package in one prompt: editor, database, server, hosting, even domain registration. If you want to one-shot a landing page, they win.
But the more Greg dug in, the more he realized Codex Sites is playing a different game.
It's not trying to be the prettiest app builder. It's trying to be the app builder that lives where your agent lives.
Greg's example is his own personal website. It says 158,000 people read his newsletter. It's a static site, so when that number hits 160,000, he has to go edit it by hand.
With Codex Sites, the agent updates it for him. New subscriber count, new guides generated from his content, new companies added to his portfolio. Automatically.
That's the actual pitch. Forget "build an app fast." This is "build an app that Codex keeps operating for you."
If you've read my piece on why AI agents are the new SaaS, this is that thesis playing out in real time.
What he built: a Startup Ideas OS in 6 prompts
Greg didn't build a demo to-do list. He built something he'd actually use: a live board for triaging startup ideas.
Columns: Inbox → Researching → Validating → Building → Killed.
Each card holds the idea, the buyer, the pain, the proof, the next step, and a score.
Six prompts, start to finish. And each prompt teaches you one concept you'll reuse on everything you build with this thing.
Prompt 1: Invoke it right, and never let it deploy early
Codex Sites works as a plugin. You have to invoke it with "@sites" or it won't kick in.
Greg's first prompt was simple: "Build a startup OS."
But he flagged two tips that save you pain:
→ Ask for realistic sample data. Otherwise you get lorem ipsum garbage you can't evaluate.
→ Say "save for review, do not deploy." Left alone, Codex will try to ship your half-built idea straight to a live URL.
One more thing most people skip: the plugins section. Figma, Canva, HeyGen for avatar videos, Remotion, FAL for image generation, Hugging Face models. All wired in from the start.
Greg's underrated pick is Game Studio. His logic: nobody visits your vibe-coded app because nobody knows it exists. But a fun little game that newsjacks a trending topic? That gets shared. Then you funnel those players into your newsletter or core product.
Prompt 2: Add memory, or you've built a screenshot
Out of the box, a Codex Sites app forgets everything between visits.
Greg's second prompt: "Add persistent storage so ideas stay saved between visits. Before coding, show me the data model and which records and actions the app needs."
That second sentence is the smart part.
Instead of letting the agent silently pick an architecture, he made it explain itself first. Codex came back with a plan: Cloudflare D1 as the durable store, one main record type (ideas), and a list of the actions the app needs. List, add, update, move, score, archive.
Without memory it's a demo. With memory it's a product.
Prompt 3: Safe actions, the real unlock
This is the concept that made me sit up.
A safe action is a named, approved operation the agent is allowed to run. "Add idea." "Move card." "Score idea." Nothing else.
The agent can't run raw SQL, can't write arbitrary rows, can't deploy behind your back. It gets a short menu of buttons and that's it.
Why does that matter? Because it means you can be in a completely different chat, ping-ponging startup ideas with Codex, hit a good one, and just say "add idea." It lands on your live board. You never opened the app.
And Greg's tip for non-technical founders is sneaky good: you probably don't know which safe actions your app needs. So don't guess. Ask Codex to tell you (that's exactly what prompt 2 did), then tell it to build those.
Prompt 4: Create a skill so future chats know how to drive the app
A skill is a reusable instruction manual for your agent.
Greg's prompt: "Create a Codex skill called Startup Ideas Admin. It should explain how to read the board, how to add ideas, how to move cards, how to score ideas, and include five example commands."
Now every future chat knows how to operate the board. He never has to re-explain it.
Greg says most people using Codex Sites aren't creating skills at all, and he thinks it's one of the best parts of the whole system. I agree. It's the difference between a tool you use and a tool your agent uses.
(If this whole workflow feels foreign, start with my breakdown of how to become AI native first. Same channel, same author, and it's the foundation this video builds on.)
Prompt 5: Save gates, because Codex doesn't auto-save
Greg's analogy: video game checkpoints.
Codex doesn't auto-save your app versions. So before anything goes near a live URL, he prompts: "Save this as V1 review. Do not deploy. Confirm build status, storage choice, access setting, and the exact version I should review."
Codex came back: build passed, storage still D1, access custom, version saved. Nothing deployed.
Boring? Yes. Also the thing that saves you when the agent goes sideways on prompt 14 and you need to roll back.
Prompt 6: Prove the loop
The final test. Greg opened a NEW chat (that part matters) and typed:
"Use Startup Ideas Admin to add this startup idea: AI agent SEO grader for local businesses. Put it in Inbox with a first-pass score and a next step."
Codex loaded the skill, read the board, called the safe action, verified the write, and reported back: "Done, using only the safe board API. No raw SQL, no generic database writes, no deploy."
The idea appeared on the live board. From a different conversation. Without Greg touching the app.
That's the loop: an app your agent operates on your behalf. Add a weekly automation and the board fills itself while you sleep.
The honest limitations (Greg names them, so will I)
Codex Sites is missing real pieces today:
→ No payments. No email sending. No analytics. No secrets vault. Auth is basic.
→ You can't publish to your own domain yet. Apps live on an ugly generated URL and are really internal tools you share with your team.
→ It feels more technical than Lovable or Bolt. You have to prompt for memory, prompt for safe actions, prompt for skills. Nothing is hand-held.
Greg's prediction is that custom domains and easier defaults are coming soon. Maybe. Plan for what exists today: internal tools, personal dashboards, and ops software for your own business. Not your customer-facing SaaS. Not yet.
How I'd use this as a bootstrapped founder
Forget the personal-website use case. Here's where this hits for someone doing $5K to $50K MRR:
→ A churn watchtower. Board of at-risk accounts, agent moves cards as usage drops, safe action to draft the save email.
→ A content pipeline OS. Ideas in, drafts researching, published tracked. Your agent adds ideas from every brainstorm chat.
→ A competitor tracker that updates itself weekly on a cron.
The pattern is always the same one Greg proved: memory + safe actions + a skill + one automation = software that runs without you.
You're not building apps anymore. You're hiring them.
FAQ
What is Codex Sites?
Codex Sites is a plugin inside OpenAI's Codex that builds and hosts web apps directly from chat. Unlike Replit or Lovable, its apps can be operated autonomously by Codex itself: the agent updates content, adds records, and runs automations through approved "safe actions" without you opening the app.
How is Codex Sites different from Replit or Lovable?
Replit and Lovable bundle everything (editor, database, hosting, domains) for one-prompt building. Codex Sites trades that polish for agent integration: apps live where your AI works, so any chat can read or update them. It's weaker for public products, stronger for internal tools your agent maintains.
What are safe actions in Codex Sites?
Safe actions are named, pre-approved operations (like "add idea" or "move card") that are the ONLY way an agent can modify your app. No raw SQL, no arbitrary writes, no surprise deploys. They're what make it safe to let an agent operate your software from any conversation.
Can I put a Codex Sites app on my own domain?
Not yet. As of Greg's video, apps deploy to a generated URL and work as internal tools you share with your team. Greg predicts custom domains are coming, but today you should build ops tools and dashboards, not customer-facing products.
Do I need to be technical to use Codex Sites?
More than for Lovable, less than for real coding. You won't write code, but you do need to prompt for storage, safe actions, and skills yourself. Greg's workaround for non-technical founders: ask Codex what data model and actions your app needs before building, then tell it to build exactly that.
Steal playbooks like this every week
I break down what actually works for bootstrapped founders on the Profitable Founder Podcast: real numbers, real operators, no fluff.
