Turn code and logs into a useful diagnosis

Start with one failure you can describe precisely. ChatGPT is useful here when the relevant evidence fits into a few files and you want to explore the cause before opening a larger coding task. This example uses synthetic JavaScript; replace its files with material you are allowed to share.

A ChatGPT project can keep sources and instructions available across related chats. For one isolated diagnosis, a new chat is enough. An uploaded file does not grant access to the rest of your checkout.

Give it a failure it can investigate

Suppose a retry helper waits too long on its first retry:

export function retryDelay(attempt) {
  return Math.min(1000 * 2 ** attempt, 30000);
}

The caller passes 1 for the first retry. A log records retry=1 delay=2000. The requirement says the delays should begin at one second and cap at thirty.

Provide the function, the caller's convention, and the observation together:

Diagnose why the first retry waits two seconds. Use the attached function, the caller convention, and the log. Identify the expression causing it, suggest the smallest correction, and give checks for attempts 1, 2, and 6. Separate evidence from assumptions. Leave the caller's numbering unchanged. If you cannot execute the code, give me commands I can run locally.

The useful answer points to the exponent and proposes attempt - 1. It also asks what should happen for zero or negative attempts, since the supplied requirement says nothing about them. That is an unresolved contract question.

Make the result usable in your editor

Request a short handoff containing the failure, relevant file and line, proposed patch, expected checks, and open questions. Keep the original files with it. A later agent should be able to inspect the evidence behind the diagnosis instead of inheriting an unexplained conclusion.

If your ChatGPT setup provides execution tools, ask it to run the example and return the command and result. Otherwise, run the checks in your development environment. An assistant saying a test would pass is a prediction.

For a task that needs broader repository access, continue with Codex's checked-change walkthrough. Pass the evidence packet and acceptance criteria, then let Codex inspect the current checkout before implementing.

What goes wrong

A log without the caller convention leaves two possible fixes: change the formula or change the numbering. A screenshot without the source makes the diagnosis harder to verify. A confident patch can also alter the cap or retry count even though the task concerns only the delay.

Keep the requested change small. Treat unobserved behavior as a question, especially where it affects an API contract.

How to check

Run the corrected function with attempts 1, 2, and 6. Expect 1000, 2000, and 30000 milliseconds. Confirm that attempt 1 on the original function returns 2000, so the check actually detects the reported failure.

Inspect the diff for changes to numbering, caps, or retry count. Resolve the invalid-input contract before adding tests that silently define it.