Skip to main content
C//CCREATIVE
COMMERCE
AI & PROMPTS

ChatGPT, Work or Codex? Choose the right AI mode for the task

A practical split between conversation, multi-step work across sources and repository-level coding with tests.

← Back to Knowledge2 min read
Prepared by Creative Commerce by PlisnerUpdated: 2026-09-19Practical content based on source material and the real service scope.

Chat, Work and Codex are not three names for the same workflow. The simplest choice is: conversation → Chat, multi-step finished result → Work, repository and code → Codex.

Chat

Best for quick conversation, analysis, drafts and a single image.

Work

For multi-step tasks, research and a finished result.

Codex

For software: repositories, files, tests and fixes.

A simple selection rule

TaskModeWhy
Write three CTA variantsChatShort result and rapid iteration.
Research the market and create a reportWorkMultiple sources and stages.
Fix a plugin and run testsCodexCode context and execution are required.

Ready commands

CHAT

Chat: quick decision

I have three variants of [X]. Compare them by cost, time, risk and user impact. Give me a table first, then a recommendation and the conditions under which you would choose a different option. Do not guess prices if I did not provide them.
WORK

Work: from research to result

Research [topic] using current, credible sources. Separate facts from opinions. Compare sources and show discrepancies and dates. Then prepare a finished [report/plan/article] with recommendations. At the end, add a list of claims that could not be confirmed unambiguously.
CODEX / CODE

Change existing code without causing regression

First audit the existing repository and identify the files responsible for the problem. Do not change the architecture unless necessary. Task: [describe the exact bug]. Preserve working APIs, routing and styles outside the scope. Make the smallest safe change. Add or update a regression test. Run lint and the full existing test suite. At the end, report: changed files, root cause, test results and anything that could not be verified. Do not mark the release as ready if the tests are red.

A code prompt should resemble a good technical ticket: current state, expected result, constraints and an acceptance test.

NEXT

Keep going

Choose a related guide or return to the Knowledge magazine.

How to create your own ChatGPT Pet: what it does and how to design a custom lookOpen →How to customize ChatGPT with Custom Instructions for more consistent workOpen →
All guides →