OpenAI now lets people bring selected projects, chats, skills, plugins and settings from supported agent tools into ChatGPT Work and Codex. The launch demo shows the move in the ChatGPT desktop app, and OpenAI’s import guide explains what transfers and what needs review. The five-year-old explanation
Think about moving a busy employee to a new desk. Their folders, checklists and recent notes contain the knowledge that makes them productive. OpenAI built a mover for several popular agent tools. The importer labels those folders for ChatGPT Work or Codex, then leaves the original desk intact.
A company that uses several AI systems needs a second layer: a small, portable handoff pack. This pack tells any capable AI what the project is for, which material is trusted, how the work gets done and what a passing result looks like. You can upload it to ChatGPT, Claude, Gemini, Copilot, Cursor or another approved system even when that product has no automatic importer.
What OpenAI can move today
ChatGPT desktop: Open Settings, choose Import, select Claude Code, Claude Cowork or Cursor, then choose the projects, chats, skills, plugins and settings you want. Import history and an automatic-updates switch live in the same area.
Codex CLI: Run /import to bring work from Claude Code or Cursor. The CLI can bring in as many as 50 chats from the previous 30 days. It cannot start an import while another task, remote session or local app-server daemon is running.
OpenAI maps source instructions into project instructions, brings supported skills and plugins across, and turns existing folders into projects. Permissions, connector sign-ins, hooks, path names and plugin behavior still need a person to review them.
Build the portable pack
1. One-page brief. State the business goal, the audience, the owner, the required inputs and the final deliverable. A marketing example might be “turn one approved webinar transcript into a newsletter, three social posts and a sales follow-up.”
2. Trusted source folder. Add the current brand guide, product facts, price sheet, policy, template and approved examples. Put dates and owners on files so the AI can spot stale material.
3. Playbook. Write the routine as short steps. Name the tool that performs each step, the fields it may change and every point where a person must approve an action.
4. Quality gate. List pass-or-fail checks for accuracy, tone, privacy, accessibility and required approvals. A clear gate gives every AI system the same finish line.
5. Sample and change log. Include one approved finished example and a short record of rule changes. The sample shows the target; the log tells the next AI which instruction is current.
Deploy it across a mixed tool stack
Start with one safe project. Choose a repeatable job that uses public or synthetic information, such as a weekly competitor summary or a campaign brief built from approved materials.
Use the automatic route where it exists. In ChatGPT desktop, import only the items needed for that test. Open the imported project and ask it to name the instructions, sources, skills and plugins it can see.
Use the pack everywhere else. Upload the same brief, sources, playbook, quality gate and sample to each approved AI product. Save the project instructions inside that tool if it supports them.
Reconnect permissions carefully. Sign in to connectors again, give each one the smallest useful level of access and place sending, publishing, purchasing, deleting and customer-facing changes behind approval.
Run the same test case. Give two systems identical inputs and compare the results against the quality gate. Fix the handoff pack when both systems make the same mistake; tune one tool when only that tool misses.
Choose one source of truth. Turn on OpenAI’s automatic updates only after your team decides which source setup owns future changes. The feature can carry supported changes into ChatGPT. It does not create universal, two-way synchronization among every AI vendor.
Copy this handoff request
Turn this project into a portable AI handoff pack. Include: purpose, owner, required inputs, trusted sources, brand and policy rules, repeatable steps, approved tools, fields each tool may change, human approval points, pass-or-fail quality checks, one sample output and known failure modes.
Remove passwords, API keys, customer records, personal data and one-time details. Label anything that may not transfer cleanly to another AI system. End with a five-minute test that confirms a new system understood the pack.
Where this helps a business
Marketing: carry voice rules, approved claims, campaign calendars and review checklists across a writing tool, an image tool and the system that schedules content.
Client service: give every approved assistant the same intake questions, account facts, escalation rules and definition of a finished client update.
Operations: document a recurring report, including where data comes from, which calculations are fixed, who checks exceptions and who sends the final version.
Software and IT: move coding instructions, repository rules, reusable skills and test expectations while rechecking paths, permissions and external services on the new system.
The guardrails
Keep passwords, API keys, financial account details, health information and identifiable customer data outside the pack. Review imported permissions and tool restrictions before a live run. Reauthorize connectors one at a time. Keep every consequential action behind a person until the imported project passes a real test with safe data.
|
