Codex: From Request to Verifiable Result

Capstone

Deliver one command-to-receipt system

Choose a real workflow you understand and use Codex to deliver one complete vertical slice with enough evidence, controls, documentation, and recovery that a second person can verify and operate it without reading the original conversation.

What you must build

• ◆A task contract with one user outcome, bounded scope, non-goals, business rules, examples, negative cases, and stop conditions.
• ◆A current repository and environment map distinguishing source, local runtime, cloud task, preview, deployment, and customer state where relevant.
• ◆A concise AGENTS.md hierarchy and control-surface map containing only durable, correctly scoped guidance.
• ◆Separate read-only investigation, local implementation, verification, and consequential release authority.
• ◆An isolated branch or worktree with a narrow diff and no unrelated user changes.
• ◆Bounded subagent or extension use only where it materially improves independence, repeatability, access, or enforcement.
• ◆Risk-matched automated checks plus the actual critical user flow and at least one meaningful negative case.
• ◆An exact release receipt when deployment is authorized, or an explicit statement that the work is not deployed.
• ◆A tested rollback or recovery procedure and an operator-readable handoff.

How you prove it

• The final result can be explained from user outcome through interface, logic, state, identity, integrations, and proof.
• Every consequential action is within explicit authority and every credential remains in an approved secret mechanism.
• The final diff and all verification commands are independently reviewable.
• The highest verified environment layer is named without overstating production or customer proof.
• A second reviewer can reproduce the core result and rollback from the delivery packet alone.
• The retrospective identifies which workflow, if any, has earned another supervised run before automation or broader autonomy.