Independent guidance from Blue Canvas. Not affiliated with or endorsed by OpenAI.

Business workflow guide

ChatGPT dots and Codex: brief a coding task for review

By Blue Canvas ยท Published and checked 16 October 2026

AI-assisted guidance prepared for Blue Canvas from the official sources linked below. The examples are proposed tests, not reports of firsthand product testing.

Use your dot to coordinate the responsibility and Codex to work on the code. The handover should identify the exact project, the evidence of the problem and what counts as a reviewed fix, without quietly authorising a production release.

What dots can hand to Codex

OpenAI's October update says a dot can start Codex work and follow up on existing threads. That is useful when a responsibility produces coding work, but it does not mean every follow-up needs another new task.

For a small business site, a recurring feedback review might identify a reproducible layout problem. Give the dot the specific issue and ask whether an existing task already covers it. Our recommendation is to keep related work together where possible and reserve a new task for genuinely separate scope.

Product source: OpenAI's official documentation, checked 16 October 2026.

Choose the execution environment deliberately

OpenAI documents local Codex tasks on a connected computer and cloud coding tasks in a Codex environment you have already configured. The local computer must stay online with the ChatGPT app open. Selecting a different computer does not move an existing task there.

Before work starts, identify the repository and the intended checkout or cloud environment. A project name alone may be ambiguous. Do not assume a similarly named demo, old copy or recovery host is the production source. Ask for the verified target before permitting changes.

Product source: OpenAI's official documentation, checked 16 October 2026.

Give evidence of the problem, not just a desired result

A useful issue report says what you did, what happened, what you expected and where the difference can be reproduced. Include the relevant page or component and the device conditions if they matter. Remove account details or personal information from screenshots and sample data.

For example, 'the heading overlaps the next section at a narrow width' is more actionable than 'make mobile better'. Ask the task to reproduce the problem first and preserve behaviour outside it. If the evidence does not reproduce, request a diagnosis rather than an unrelated redesign. This is a proposed workflow, not a claim about a completed customer fix.

A review-only coding handover

Adjust this brief to your team's development rules. A repository may require a particular branch, worktree, review or release process. Do not ask the agent to bypass a failing check just to produce a green status; have it distinguish an actual failure from a check that could not run.

For an existing task, name that task and the exact follow-up instead of repeating the whole request as though it were new. Ask what has changed since the previous review so the reviewer is not comparing two unrelated versions.

Adapt this example to your approved task

Project: [exact repository and approved environment].
Problem: [reproduction steps, expected result and observed result].
Scope: investigate and fix only [named behaviour]. Preserve unrelated work.
Inputs: [approved issue and sanitised fixture].
Checks: run the project's applicable tests; verify the original problem and relevant mobile/desktop states. Report anything not tested.
Boundary: prepare changes for review. Do not push, merge, deploy, contact users or submit real forms. Do not read secrets or use live customer data.
Return: changed files, source revision, test evidence, screenshots where relevant, risks and the next decision needed.

Review the actual change, not only the summary

OpenAI's controls documentation points to Activity in the desktop app for inspecting delegated progress, files, results and requests for input. Use the task's evidence to decide what to review next; a confident completion message is not a substitute for a working result.

Our suggested review has three parts. First, check that the change belongs to the intended repository and addresses the reported issue. Then inspect the tests and rendered result, including any failure or untested path. Finally, look for unrelated changes to data handling, permissions, payments or navigation that were not part of the brief.

Product source: OpenAI's official documentation, checked 16 October 2026.

Keep preparation and release as separate decisions

Record the revision you reviewed, who can authorise release and the destination to be updated. A local test result supports that local revision; it does not prove what a live domain serves. After an authorised release, the release owner should check the deployed revision and the real user journey without creating real enquiries or transactions during testing.

If the source changes after review, inspect the difference and rerun the relevant checks before release. If a dependency or connection is missing, keep the result labelled prepared or blocked rather than published. This discipline is useful whether the work was coordinated by a dot, a person or an existing issue tracker.