ChatGPT Projects in Codex: How to Hand Off the Work
Move the files, instructions, and approved conclusions a coding task needs. Check the working folder and current state before asking Codex to make changes.
Updated 5 min read
To use ChatGPT project work in Codex, hand over the relevant files, instructions, approved decisions, and next task explicitly. OpenAI's project documentation (opens in a new tab) says the CLI and IDE extension use local folders rather than exposing the ChatGPT Projects view; sharing a sign-in does not establish which materials a session can read.
A useful handoff makes the project's current state clear before code changes begin. fenbs can hold the shared tasks, rules, and Where we left off so the work record remains available when you move between connected ChatGPT and Codex sessions.
What should you do before opening the coding task?
- Choose the exact outcome. Describe one change and how you will recognize a correct result.
- Collect the current project files from their original locations, including any reviewed outputs that exist only in the ChatGPT project.
- Copy the relevant project instructions into a reviewed handoff, separating permanent repository rules from this task's requirements.
- Open the intended folder or coding environment. Ask Codex to identify the files it can read and summarize the work before editing.
- Compare that summary with your handoff, then authorize the next bounded step.
For example, turn “continue the website project” into “add the approved contact-form fields, preserve the existing delivery service, and show which validation checks passed.” The second request gives a coding assistant a result to work toward and gives you something concrete to review.
Keep environment setup separate from project knowledge. OpenAI's quickstart (opens in a new tab) explains the available Codex surfaces and selecting where work runs. Our ChatGPT Codex overview provides the broader product context; the task here is transferring the material for one continuing project.
Which parts of the ChatGPT project matter?
Make an inventory rather than trusting a final chat summary to contain everything. OpenAI describes ChatGPT projects (opens in a new tab) as bringing chats, files, and instructions together. Review each kind because each can contain something the coding task needs.
| Material | What to carry forward | What to review |
|---|---|---|
| Instructions | The task requirements and rules that apply to the code. | Remove advice that only applied to an earlier writing or planning session. |
| Files | The current source files, assets, and relevant reference documents. | Check names and versions against the actual working folder. |
| Conclusions | Approved choices and the reason behind each choice. | Separate a person's approval from an assistant's suggested approach. |
| Progress | What changed, what was checked, and the next unresolved step. | Distinguish working code from an untested proposal. |
A source document and the interpretation of that document serve different purposes. Carry the approved requirements as well as the reasoning that affects the implementation. If the interpretation conflicts with the source, ask a person to settle it before turning it into code.
The project-management guide covers organizing work inside ChatGPT. Here, the test is narrower: does the destination coding session have the material needed to make the next change correctly?
Where should durable instructions live?
For a repository, put reviewed engineering guidance where future coding sessions can find it. OpenAI documents that Codex reads AGENTS.md instructions (opens in a new tab) before working, with guidance layered by directory. Follow the repository's existing conventions rather than creating competing instruction files.
Keep the durable file focused on commands, code conventions, and rules that repeatedly apply. Put the immediate deliverable, temporary investigation notes, and current test results in the task handoff. A rule like “do not change the payment flow” needs a clear scope and a reason.
Do not turn every suggestion from the old conversation into a standing rule. “We could change the database later” is not permission to change it now. Keep unanswered questions explicit, especially when they affect customer behavior, data, or an external service.
fenbs keeps people's Decisions and rules alongside the task, so a connected assistant can read the business choice behind repository guidance. Its handoff workflow also preserves the current next step when that guidance stays the same.
How do you check that Codex has the right context?
Start with a read-only request. Ask the assistant to identify the working folder, locate the relevant files, and explain the intended behavior. Ask what evidence it still lacks. This gives you a chance to catch a wrong checkout or a missing asset before changes spread.
- Does the session see the current source files rather than a copied draft?
- Can it name the approved requirement and the person's unresolved question?
- Does it know which checks actually ran and which remain?
- Is the requested change small enough to review separately?
If the answers are incomplete, supply the missing evidence. A private link is only useful when the assistant can actually open it. A filename without the file is not a transferred document. Request a revised plan once the missing context is available.
What causes a handoff to go wrong?
The most common practical mistake is moving conclusions without their conditions. “The form works” may mean someone reviewed the layout, while delivery was never tested. Write the exact check and result, then state the remaining limitation in ordinary words.
Another mistake is asking the new session to recreate missing context by guessing. If an approved design is unavailable, pause that part of the change. The assistant can inspect the code and list questions while you recover the source, but the missing design should not silently become an invented requirement.
Keep the broader migration separate: the guide to switching away from ChatGPT covers choosing and preparing a destination. For this handoff, retain the original project until the new session has successfully completed a small, reviewed task.
Carry the project state into the coding session
Record the next coding task, its plan, the approved choices, and unfinished checks in fenbs. Connected Codex sessions can read the same context as ChatGPT and save progress back to it. The source code stays in its repository; fenbs holds the work record that explains what to do next.
Create a free accountFree for one person · No credit card · Works with Claude, ChatGPT, Cursor and more
How to do this with fenbs
- Create a free account and add the project that will continue in Codex.
- Add the coding task with its goal and plan. Record agreed instructions in Decisions and rules, and leave unanswered choices open.
- Connect ChatGPT and Codex using their integration instructions. MCP is a standard connection that lets an assistant read and update an outside tool.
- Before the coding session, ask it to read the project context and state the next step. Afterward, ask it to save what changed, the checks run, and what remains.
fenbs does not import ChatGPT projects or copy your repository. It is unnecessary when a single small prompt and a local file already explain the whole task. Use it when the project will continue across sessions and its decisions need a durable shared home.
For more ways to keep a project clear before the handoff, read ChatGPT project best practices. The switching models collection covers keeping useful work available across other assistants too.