Lesson 03 / 06 · Codex
Build with Codex and skills.
Move from an attractive prototype to a tested change in the owning repository.
Your artifact: A bounded patch, checks and preview
Give the builder acceptance evidence.
Codex needs the actual source, the intended outcome and the constraints that define a good change. Read AGENTS.md and existing patterns before choosing a new dependency. Supply the visual target alongside the behavior it represents. Keep an approved prototype and the production implementation tied to explicit revisions.
A reusable skill packages a repeatable method and its resources. A plugin can bundle skills and tool connections. Neither grants permissions by existing as a Markdown file. Reuse Starlight's installed architecture and design capabilities when available, then verify the real host binding instead of inventing an @name.
Ask for a vertical slice: a working input, a meaningful result, recovery and export. Checks must inspect behavior and failure states. A successful build establishes compilation; a preview establishes a hosted candidate. Production and user usefulness need their own evidence.
Worked example · illustrative
See the decision.
A comparison prototype becomes a server-rendered course page with a small client workbench. The repository already provides navigation, security headers and typography. Reusing them avoids a competing application shell and makes the new learning surface discoverable.
Run the prompt.
Use the supported host shown above and supply your own non-confidential context. Preview availability and tool access depend on your account.
Implement this reviewed visual artifact in the repository that owns the production route. Read applicable AGENTS.md, architecture decisions, dependencies and release instructions. Record the exact base revision. Reuse installed Starlight AI Architect and relevant design/verification skills when present; report unavailable bindings. Preserve unrelated changes. Implement one complete interaction with accessible narrow-screen, keyboard, error and export behavior. Run the required repository checks and inspect a Git-backed preview. Return the patch, actual check results, preview source SHA, rollback and unresolved limitations. Release only within my authorized scope and the repository's acceptance gates.Open the prompt library and studio
Make a first attempt.
Give Codex an existing repository and your prototype. Before implementation, freeze three observable acceptance statements and one failure case. Compare the final preview against them.
Inspect your artifact
- The owning repository and exact base revision are recorded.
- Existing identity, billing and runtime boundaries remain coherent.
- Checks cover the chosen interaction and at least one failure.
- The preview is tied to the actual candidate source.
Transfer / A fresh case
Carry the idea further.
Package only the repeatable part as a skill proposal. Test it on a second artifact before calling it reusable.
Keep a note of your prediction, what happened and one revision. These are self-study checks, not a graded assessment.
Read the primary sources.
Checked 2026-10-20. Recheck availability, access and API versions before a real implementation.
- OpenAI: Custom instructions with AGENTS.md
Repository guidance before implementation.
- OpenAI: Build skills
Reusable instructions and supporting resources; installation and tool authority are separate.
- OpenAI: ChatGPT Work and Codex
Choose a surface by the task; account and rollout access vary.