Your code. A little further.

Start with the right project context

• Attach a repository or relevant files

  Use the repository attachment workflow to load the project you want to discuss. Inspect the repository, branch, revision and file preview before attaching it. A connected GitHub account is not proof that every repository or file has been read. Keep credentials and private environment files out of chat.

• Ask for a map you can verify

  Begin with the entry point, dependency manifest and the files involved in one feature. Ask Baby 5.0 to name the files it inspected and identify anything it cannot access. This gives you a better starting point than a broad summary that could have been written from the README alone.

• Keep one outcome in scope

  For an existing app, name the behavior you want to change and the behavior that must stay the same. For a new app, describe the first useful screen and what should happen after a user action. Save a working version before adding more features.

Debug a reproducible problem

• Describe the trigger and the result

  Include the steps that reproduce the issue, the expected result, the actual result and a relevant error message. For example: a task appears after saving, but disappears after reloading the page. That is a clearer problem than asking the assistant to fix everything.

• Request a bounded investigation

  Ask the assistant to trace the save action through the form, request and persistence code. Request file-backed findings and a small proposed change. If a required file or tool is unavailable, the assistant should identify that gap instead of claiming it ran a check.

• Verify the proposed fix

  Run the checks in your actual development environment and repeat the failing action. Inspect the stored result and try the surrounding workflow. Generating a test or suggesting a command does not establish that it executed or passed. Keep the output needed to explain the result to a reviewer.

Worked example: debug JavaScript numeric sorting

• Reproduce the failure

  For finite numbers, [2, 10, 1].sort() returns [1, 10, 2], rather than the expected [1, 2, 10]. JavaScript’s default sort compares string representations and changes the original array. An AI coding review should identify both the ordering issue and the mutation before proposing a change.

• Apply a bounded fix

  Copy the array and provide a numeric comparator: function ascendingNumbers(values) { return [...values].sort((a, b) => a - b); } This example accepts finite numbers. Decide separately how your application should handle numeric strings, missing values or NaN.

• Verify the result yourself

  The downloadable example checks ordinary numbers, an empty array, duplicate values, negative numbers and preservation of the input. Run it with Node.js: node numeric-sort-debugging.mjs. The example is a reproducible fixture; it is not proof that an AI-generated change in your own repository passes its tests.

A fictional task-manager brief

• Use a testable starting request

  Example brief: Create a task manager with a task list, an add-task form and a completion control. Require a nonempty task name, show a useful empty state, and keep tasks after closing and reopening the app. This is an illustrative brief, not a claim that a production application has already been generated or tested.

• Review the first iteration

  Add an empty name, a long name and several tasks in succession. Complete a task, navigate away and return. Reload the app and check persistence. On a phone, check the keyboard, small-screen layout and touch controls. A convincing screen can still have a placeholder save action.

• Choose a targeted follow-up

  If the first result loses tasks after a reload, describe that specific failure and ask for persistence to be corrected. Retest before adding sorting, categories or account sharing. Smaller iterations make it easier to identify the change that caused a regression and to keep a usable version.

Choose the workflow that fits

• Understand or revise an existing codebase

  Use chat with the relevant repository context to explain a feature, investigate a failure or review a proposed change. Access and available actions depend on the connected account, project and tools. Do not assume a chat session includes a runnable environment or permission to publish code.

• Build a mobile or web prototype

  The mobile builder starts a React Native and Expo project for a phone workflow. The web builder starts a website or web application. Choose the destination before building, then review the source, configuration and behavior. A prompt does not complete signing, hosting or store review on your behalf.

• Review cost and evidence

  Check current plans, shared builder usage and account allowances before a larger project. Assess the complete process: correctness, necessary revisions, usage and the work still needed for release. This page does not claim a fixed context limit, a training corpus, a benchmark score or a cost advantage over another model.

Frequently asked questions