A useful handoff from ChatGPT to Codex is a small set of accepted product decisions and implementation criteria. Give the coding task the context it needs to build and verify one outcome, rather than asking it to recover your business intent from a conversation transcript.
This guide describes a manual file-based handoff. It does not require an automatic bridge between 1PC and your repository.
For the broader context-maintenance method, read how to share product context across AI tools. It covers saved knowledge and connection checks; this guide keeps the implementation handoff manual while Codex-specific 1PC verification remains outstanding.
Handoff a decision, not a transcript
Product discovery includes ideas you rejected, claims you have not tested, and details you later changed. Exporting everything can give a coding assistant conflicting instructions.
Before the handoff, select the current customer, problem, evidence, scope, constraints, and acceptance criteria. Flag unresolved decisions. A vague requirement should stay unresolved until you can decide it, not become an accidental implementation choice.
Use the MVP planning guide to draw that boundary. The context guide explains how to keep product facts current.
Prepare a small handoff package
These are suggested repository paths for your own project, not built-in 1PC exports:
docs/product-context.md
Current customer, problem, evidence, founder constraints.
docs/mvp-scope.md
One outcome, included behavior, and excluded work.
docs/acceptance-criteria.md
Observable success, errors, access boundaries, and checks.
AGENTS.md
Repository conventions and instructions for coding work.
If you keep files in 1PC, download the relevant current documents and review them before copying into your repository. Remove information the implementation does not need. Confirm the dates and resolve contradictions.
Keep company context separate from repository instructions. Product documents explain why and what; repository guidance explains how to work in that codebase. The official OpenAI project guide describes checked-in guidance as a way to make durable project instructions available to future work.
Write a scoped implementation prompt
For the hypothetical AgentCRM pilot, a handoff might look like this:
Read the repository instructions and these accepted documents:
- docs/product-context.md
- docs/mvp-scope.md
- docs/acceptance-criteria.md
Implement the inquiry next-action workflow in the existing stack.
Inspect current authentication and record ownership first.
The release includes:
- adding an inquiry with a next-action date;
- listing the signed-in agent's actions;
- marking an action complete;
- useful empty and invalid-input states.
Do not add campaigns, integrations, or team administration.
If a business rule is unresolved, identify it before coding.
Verify the user journey and account ownership boundaries.
Report what changed, checks run, and remaining limitations.
The prompt is effective because the documents already make the product choice reviewable. It is not a substitute for inspecting the existing code or resolving unknown requirements.
Verify the user outcome
Review the implementation against acceptance criteria, not only the diff size. Try the first-use state, valid task completion, invalid input, a failed save, and access from another account. Check the actual rendered interface on mobile when the user will work there.
Ask for relevant automated checks and a clear explanation of what they establish. A passing unit test does not establish that the complete onboarding flow is usable. A screenshot does not establish account isolation.
Keep deployments and irreversible data changes as explicit decisions. A reviewed build brief authorizes implementation work, but your release process still needs its own concrete checks.
Bring the learning back into product context
When the pilot runs, update the overview with implemented behavior and limitations. Save the user observations that change the next decision. Do not mark a feature as available merely because it was discussed or added to a roadmap.
In 1PC, current knowledge files can be edited and downloaded. The MVP does not automatically synchronize repository changes or provide a dedicated Codex bridge. A manual loop can still be useful: accepted context → build brief → reviewed implementation → updated context.
Return to the founder workflow to connect the next customer experiment to what you actually built.
Prepare the context before the code.
Use a compact founder context file to keep your build brief aligned with your actual resources.
Get the context template ↗