OpenAI Codex: Build Apps That Work For You 24/7
TL;DR
Codex Sites is about autonomous app updates, not one-shot site generation: Greg says Replit and Lovable are stronger if you want editor, database, hosting, payments, and domains out of the box, but Codex Sites gets interesting when you want apps that Codex itself can keep operating over time.
The first prompt should force review mode and realistic data: His practical tip is to invoke the Sites plugin, ask for realistic sample data, and explicitly say "save for review, do not deploy" so Codex builds a real product workflow instead of rushing to publish a homepage.
Memory turns a demo into usable software: He adds persistent storage with Cloudflare D1 so startup ideas survive between visits, with records and actions like add idea, move idea, score idea, and archive idea.
Safe actions are the key control layer: By creating named actions like add idea and update idea, Codex can modify the app later without raw SQL or arbitrary database writes, which is what makes cross-chat automation possible.
Skills make the app reusable from future chats: Greg creates a Codex skill called Startup Ideas Admin so later prompts know how to read the board, move cards, score ideas, and run five example commands without re-explaining the app.
The wow moment is updating the live app from a new chat: He proves the loop by asking a fresh thread to add "AI agent SEO grader for local businesses" into the inbox with a score and next step, then sees the board update through the safe API.
The Breakdown
The big trick with Codex Sites is not making a prettier Lovable clone. It is building internal apps that Codex can keep updating for you through memory, safe actions, and skills, which Greg demonstrates by building a startup ideas board in six prompts and then adding a new idea from a separate chat.
Was This Useful?
Share
Keep Reading
Make Alcreon Yours
Tune your feedFive quick questions, and the feed ranks what matters to you first.Or just get notified
The weekly Echo. Signal worth keeping in your inbox.
Every new piece, announced on X.
Read Next
See all
Playbook
Cheap Models, Hard Tasks
Most agent workflows route every step to the frontier model by default. The bill scales with how chatty the agent gets, even when most steps don't need that brain.

Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.

Playbook
The Art of Tasteful Prompting
Learn how tasteful prompting helps you move beyond generic AI output by shaping context, style, and judgment from the start.