In 90 minutes, go from installing Codex to merging a reviewed pull request on a real codebase — and leave with a setup your team can keep using on Monday.
For developers and technical founders at early-stage startups. Bring a laptop; no prior Codex experience needed.
Codex changes quickly. Commands in this kit were written against the Codex CLI and docs at the time of writing. Check
codex --helpand developers.openai.com/codex before each session and update the command check in the facilitator guide.
| # | Module | Time | You leave with |
|---|---|---|---|
| 1 | Setup | 10 min | Codex running against the starter app |
| 2 | AGENTS.md |
10 min | Project instructions Codex follows without reminders |
| 3 | Feature end to end | 25 min | A real feature across domain, API and tests |
| 4 | Tests and bugs | 15 min | The habit: failing test first, then the fix |
| 5 | Review and PR | 15 min | A reviewed pull request and a clear view of where humans decide |
| 6 | Automation | 10 min | One recipe your team can run in CI |
| — | Wrap-up | 5 min | One workflow you commit to keeping |
Half-day version: after module 2, teams switch to their own repository, write their own AGENTS.md, and run modules 3–5 on a real backlog item.
- Node.js 24 or later, Git, and a GitHub account
- A ChatGPT plan that includes Codex, or an OpenAI API key
- The Codex CLI installed (see module 1); VS Code with the Codex extension is optional
git clone https://github.com/<you>/codex-startup-workshop.git
cd codex-startup-workshop/starter-app
npm install && npm teststarter-app/ Launchlist: a small, tested waitlist API with a real backlog (ISSUES.md)
modules/ participant instructions for each module
modules/reference/ a reference AGENTS.md to compare against after module 2
automations/ reusable recipes: PR review in CI, issue → draft PR, dependency upgrades
facilitator-guide.md timings, setup checklist, common blockers, fallbacks
talk/ 20-minute meetup version with a live-demo script
feedback-template.md what to capture after each session
- Real code, not toy prompts. The starter app has tests, a typecheck and a backlog with a feature, a bug and a papercut.
- Adoption over the wow demo. Every module ends with something the team keeps.
- Humans stay accountable. Review, approval and merge are explicit steps.
- Measure it. Pick one metric before you start: issue-to-merge lead time, share of PRs with agent involvement, or review rework.
Everything works without a facilitator: follow the modules in order inside starter-app/.
MIT. Contributions and session reports are welcome via issues and pull requests.