How the airline's team uses ChatGPT and Codex

What ChatGPT and Codex do for a team

ChatGPT offers Chat for an answer or a short draft, and ChatGPT Work for a task with a clear outcome, such as a brief, an analysis or a file to review.1 OpenAI's Codex CLI runs Codex in a repository from your terminal, where it inspects code, edits files and runs commands.2 At the fictional airline, Priya, the product manager, keeps the rebooking assistant's case in a ChatGPT project. Arjun, the architect, writes the file Codex reads before it works, and Maya, the QA lead, teaches its reviewer the one rule that matters most.

When to reach for something else

• When you are tuning the rebooking assistant's own system prompt. OpenAI's prompt engineering guide says to use the Playground to develop and iterate on prompts.3 OpenAI's prompting guide says to treat prompts as application code, reviewed in the same pull requests as the behaviour they support.4 So the prompt that ships lives in git, beside the tests that judge it.
• When the 500 past cases need a home for their checks. OpenAI's Evals platform is deprecated: existing evals become read-only on 31 October 2026, and its dashboard and API are scheduled to shut down on 30 November 2026.5 Keep the cases and the checks in the repository, where CI runs them.
• When a cloud task would need a browser, or the code lives on GitLab. Computer and browser use, GitLab and self-hosted GitHub Enterprise Server are not yet supported in Codex Cloud environments.6

Five moves the airline's team makes with it

Each move names who makes it, in which phase, and the words they use. Copy one and change the nouns.

1. Move 1P1Priya, product manager

   Keep the whole case in one project

   Priya opens a project for the rebooking assistant and adds the PRD-lite, the interview notes and the glossary to its sources. A ChatGPT project gives its chats access to the same uploaded files, project instructions and connected sources.7

   A ChatGPT project has no direct access to a folder on your computer, so you upload or connect the sources you want it to use.8 So each time the spec changes on her laptop, Priya uploads the new version and removes the old one.

   Project instructions Copy

   This project is SkyWays' rebooking assistant. Owner: Priya (product).
   Read the PRD in this project's sources before you answer.
   Take every number from the PRD: the 38-minute wait, the 240 cases a day, the $400 refund limit.
   Never supply a number yourself.
   Where the PRD is silent, write NOT DECIDED, the question, and who answers it.

2. Move 2P1Priya, product manager

   Chase the open questions every morning

   A spec waits on its owners, and chasing them is a daily chore. Priya hands it to a scheduled task in ChatGPT Work. On the web a scheduled task can use uploaded files, connected tools, skills and plugins, but it cannot work directly in a folder on your computer.9

   The docs advise testing a scheduled task's prompt in a regular chat first, then reviewing its first few runs.10 Priya ran the prompt by hand for a week before she scheduled it.

   Scheduled task, weekdays at 08:00 Copy

   Read the rebooking spec in @Google Drive and the #rebooking channel in @Slack.
   List every spec field still marked NOT DECIDED, its owner, and the days it has been open, oldest first.
   Do not answer any of them. Do not post anywhere. Send the list to me.

3. Move 3P1Arjun, architect

   Write the file every Codex session reads

   Codex reads AGENTS.md files before it does any work, from the repository's root down to the folder it starts in, and a nearer file overrides the ones above it.11 Arjun writes the repository's file before the first session, and reviews changes to it like code.

   If a repository has AGENTS.md and no CLAUDE.md or CLAUDE.local.md, Claude Code reads AGENTS.md.12 So the airline keeps one file, AGENTS.md, and Codex and Claude Code both read it.

   AGENTS.md Copy

   # Rebooking service
   
   The spec is docs/spec.md. Read fields 3, 5 and 7 before you change behaviour.
   The $400 refund limit is one constant in refund_tool.py. Never copy it into a prompt.
   Every limit in the spec's field 5 is enforced in code, with a test below, at and above it.
   Never book on a stale fare: when the fare engine is down, the case queues for an agent.
   Run the tests before you call a change done.

4. Move 4P2Maya, QA lead

   Teach the reviewer the one rule that matters

   For reviews on GitHub, Codex follows the ## Code Review Rules section of the AGENTS.md file closest to the code the rules govern.13 Maya writes down the rule reviewers most often have to explain, then asks for a review on the pull request.

   On GitHub, Codex flags only P0 and P1 issues, so its review comments stay on high-priority risks.14 OpenAI's guide says review rules guide Codex and do not replace tests, branch protections or required approvals.15 So the test at $400.01 is still the gate.

   AGENTS.md, then a comment on the pull request Copy

   ## Code Review Rules
   
   ### Refund limit
   Flag any change that types 400 as a number, or moves the refund limit into a prompt.
   Safe path: import REFUND_LIMIT from refund_tool.py.
   Flag any refund path without a test at $400 and at $400.01.
   
   @codex review for the refund path and its limit

5. Move 5P2An engineer

   Hand the evaluation run to Codex Cloud

   Each Codex Cloud task gets its own isolated workspace from a published environment of repositories, dependencies and tools, and keeps working while your computer is asleep.16 The run on the 500 past cases takes a while, so the engineer sends it to the environment Arjun published.

   From the terminal, codex cloud exec --env sends a task to a cloud environment, and the CLI reference marks codex cloud experimental.17 The push comes first. A cloud task works from its environment, and your computer's local files, running processes, browser sign-ins and VPN access are not carried over to it automatically.18

   Terminal Copy

   git push
   codex cloud exec --env <rebooking-env-id> \
     'Run the evaluation on the 500 past cases in eval/cases.csv. Report the share handled right for each kind of case, codeshare as its own group, against the bar in docs/spec.md field 6. Change no code.'

Where ChatGPT and Codex let you down

• Its instructions have a ceiling. Codex stops adding AGENTS.md files once their combined size reaches the project_doc_max_bytes limit, 32 KiB by default.19 Keep the root file short, and put a service's rules in that service's folder.
• The browser stays in ChatGPT. ChatGPT's browser works on the web and in the desktop app, and is not available in Codex CLI or the IDE extension.20 So the walk through the agent's screen happens in the desktop app, not the terminal.
• A cloud task is a workbench, not a record. A cloud task's saved state does not replace source control, and the docs say to commit the work that matters.21

Three settings decide how it behaves

The traps that catch a team

• A scheduled task does not stop to ask. In the desktop app a scheduled task runs unattended with your default sandbox settings, and without approval prompts wherever your organisation's policy allows.29 With full access, a scheduled task in the background may change files, run commands and use the network without asking.30
• An @codex comment on a pull request that asks for anything but a review starts a legacy cloud chat with the pull request as context, and it can push a fix to the branch when it has permission.31 Only @codex review asks for a review.
• Code review and the GitHub and Linear integrations still run on Codex Cloud (Legacy), which OpenAI plans to deprecate.32
• The flag --dangerously-bypass-approvals-and-sandbox, alias --yolo, removes both the sandbox and the approvals, and web search then defaults to live results.33
• OpenAI warns against setting OPENAI_API_KEY or CODEX_API_KEY for a whole CI job that runs the repository's code, since its build scripts, tests and dependency hooks can read it.34
• The approval policy untrusted is retired, and a configuration that still sets it can stop Codex or ChatGPT Work from starting.35
• Prompts kept in OpenAI's dashboard have an end date. Reusable prompt objects in OpenAI's API are deprecated, and the v1/prompts API is scheduled to shut down on 30 November 2026.36

Where every fact on this page comes from

Each numbered mark above points to a row here: the fact, the vendor's page it was checked against, and the date. A date turns amber once it is more than 60 days old. Prices, usage limits and model names change too often to print: read them on ChatGPT's plans and prices, the Work and Codex pricing page and the models page.