Building a dev loop inside Codex with hooks and skills — Codex Meetup Seoul talk

In July 2026 I gave a short talk called "Codex, your way" at OpenAI Codex Meetup Seoul. The point was simple: a Codex plugin cannot change the system prompt, yet hooks and skills alone are enough to run a plan → audit → build → check → done loop inside Codex. The result is codexclaw.

The starting point: two things were missing

I studied economics and have never worked as a developer. Early this year I built a personal agent tool called cli-jaw, and that is where I first designed a structured dev loop. When the Codex app got really good I wanted the same workflow there, but two things were missing.

• Model flexibility — solved with opencodex, a local proxy that translates the Responses API for other models.
• My own loop — solved with codexclaw.

The shape of the loop

Most of the time it is just a Codex chat. When a large task arrives, the loop switches on.

1. Plan — write a plan that names the files to change and how to check them.
2. Audit — a read-only reviewer agent checks the plan against the code for gaps, before any code is written.
3. Build — implement the plan.
4. Check — collect fresh evidence: builds, tests, the real screen.
5. Done — record what changed and what was verified, then close.

If requirements get fuzzy, the loop can go back to an interview phase at any point without losing context. You can approve each phase yourself, or let the agent carry a goal all the way through.

Why a plugin was enough

A plugin cannot touch the system prompt or add slash commands. So it uses two things only.

• Hooks — every hook payload carries a session_id. That maps the Codex session to state files under .codexclaw/, so the plugin knows the current phase and can inject guidance at session start or on every turn.
• Skills — type $ in Codex and pick one; it behaves like a custom command.

One rule makes this safe: the human moves the loop freely; the agent earns every step with evidence. No test output or build exit code, no next phase.

Build tools, not backends

The line I most wanted to land: you do not need a complex AI backend to automate your own workflow. Often a SKILL.md and a simple screen are enough — no servers, no infrastructure, just markdown and a dashboard.

This site was built the same way: a written plan first, a reviewer agent auditing it, and a build plus screenshots at every phase before moving on.

codexclaw, opencodex and cli-jaw are all open source on GitHub.