I want the first question to be what you need back. A paragraph, a spreadsheet and a tested website change can begin with the same empty message box. They need very different definitions of finished.
I call this the deliverable test. Choose the environment that makes the result easiest to produce and inspect, then worry about the model. Otherwise the first afternoon becomes a tour of buttons (a remarkably easy way to feel productive).
Start with the output
A product name tells you less than the deliverable. If you want to compare two positioning statements, an ordinary conversation may be enough. If you want a board-ready pipeline workbook with formulas and a source sheet, specify that workbook. If you want the website to calculate pipeline coverage correctly, name the repository and the behavior that must change.
I would write one sentence before opening anything: “When this is done, I will have...” That sentence exposes missing decisions. A useful account brief has an audience, a freshness requirement and evidence. “Research this account” leaves all three to interpretation. Even an excellent answer can be useless when it solves the wrong version of the task.
For the examples in this hub, Cedar Metrics is a fictional B2B software company. Its account records and campaign figures are illustrative. They demonstrate how to work and check a result; they are not customer outcomes or claims about a live deployment.
TipName the file, answer or change you expect before choosing a product.
Chat is the conversation layer
Chat is the natural place to ask questions, explore alternatives and revise a short draft. Give it the situation that changes the answer: audience, source material, constraints and the decision you are trying to make. Follow up when a claim seems vague. You do not need a ceremonial prompt with a job title, ten adjectives and a fictional Nobel Prize.
For a marketer, a useful first exchange might compare two versions of a campaign promise against the same ICP. Ask which claims are supported by the supplied material and which need evidence. The result is a conversation you can use to clarify your thinking before producing a larger deliverable.
The boundary is not a ban on tools or files. Chat can use available capabilities too. The practical distinction is how you want to work: back-and-forth exploration versus handing over an outcome with enough context to carry it through.
TipStart with the question you need answered, then add the evidence that would change the answer.
Work turns a brief into a deliverable
Work is the task-oriented experience. You describe the result, supply sources and follow the progress while it gathers information, uses supported tools and prepares output. A spreadsheet, presentation, report or recurring update is a better-shaped request than “help with marketing.”
I would hand it the fictional Cedar Metrics campaign export and ask for a workbook with raw data, calculations and a summary. Require formulas to reconcile to the input totals. That instruction gives the agent both something to make and a way to know whether it worked. The finished file remains a draft until someone inspects the numbers and conclusions.
Work and Codex have overlapping capabilities. Choosing Work favors a presentation centered on the deliverable and less implementation detail. It does not imply that Code-related operations are impossible there, or that you must switch products every time a task needs a calculation.
TipA task brief should include how you will accept the result, not just the topic.
Codex gives technical work a visible workspace
Codex is useful when you need to see and manage the machinery: repository files, terminal commands, tests and diffs. A diff shows what changed between two versions of a file. That view matters when a small-looking edit can alter website behavior or a data transformation.
A GTM operator can use this without pretending to be a software engineer. A content repository, a CSV-cleaning script and a landing-page calculator are concrete systems. Ask Codex to explain the relevant files first. Then request a bounded change and an appropriate check. Understanding the changed behavior is more important than recognizing every line of syntax.
The mistake would be treating Codex as a guarantee of correctness. A successful command proves that command completed. It does not prove the business rule was right. Review the requirement, the change and the evidence together. The review guide explains that loop.
TipUse Codex when the technical review surface helps you understand and accept the work.
The model is a separate decision
ChatGPT and Codex use models, but the product name and model name are different layers. The product supplies tools, permissions, files, history and an interface. The model reasons over the context it receives. A more capable model cannot read a source that the environment cannot access.
I would start with the default option available to the account, complete a representative task and inspect the weak point. Missing source? Fix access. Wrong calculation? Require a checked computation. Weak reasoning over conflicting evidence? Compare a more capable setting. Upgrading before diagnosing the failure is like replacing a car engine because the map has the wrong destination.
Model menus and rollout availability change. The model guide separates current Work and Codex guidance from general ChatGPT assumptions. A screenshot from another account does not establish what your workspace can select.
TipCheck the surface and sign-in method alongside the model name.
Cloud and local describe execution
A cloud task runs in a hosted environment. It can continue without your laptop when the required sources and tools are accessible there. Local work uses the environment on your computer, which matters for local folders, installed programs and device-specific tools. The screen displaying the conversation does not settle where the task executes.
Imagine the account brief needs an uploaded notes file and a connected document store. That may fit cloud work. If the job must transform a folder of local exports using your installed script, local work may be the right environment. Check the dependencies before assuming closing the laptop is harmless.
Local execution also does not mean the model runs offline. Files and tool results can still be sent for hosted processing. Keep execution location, data handling and permission to act as separate questions. That distinction is useful even for a simple weekly report.
TipAsk which files and tools the chosen environment can actually access.
The API puts your software in charge
The API is the route for software that must initiate requests, store results and handle failures. Your application decides when to call a model and what happens next. It also owns the integration work: authentication, retries, validation, monitoring and the behavior when a response cannot be used.
For Cedar Metrics, a person preparing five meeting briefs could use Work. A service that scores each newly created lead and returns validated fields to a CRM is an API-shaped problem. Volume is one consideration, but control over the workflow is the stronger reason to build.
The accounts and bills are separate. A ChatGPT subscription does not by itself establish API budget or permission for an API project. Begin with the API quickstart only when you need that integration control. The shortest route to a useful result often remains the product you already have.
TipWrite down who handles a failed request before calling a workflow production-ready.
Where beginners lose the thread
The first trap is collecting features before completing a task. A Project, a plugin and a clever model setting are valuable when each removes a specific obstacle. Installed capabilities without a named job are a drawer full of adapters. You still need to know what you are connecting.
The second is accepting confident prose without checking the evidence. Open source links, inspect generated files and verify totals. If an account fact has no source, label it unknown. If a page could not be accessed, preserve that gap. A cleanly described limitation is more useful than a plausible sentence that sends a rep in the wrong direction.
The third is moving too quickly from a draft to an external action. Creating a follow-up email and sending it are separate milestones. Changing a CRM record needs the correct account, field and new value. Define that boundary in the request, then inspect the proposed result. Which small task could you finish and verify this week?
How to set it up
Choose one outcome
Pick a one-page account brief, a campaign workbook or a specific repository change. Use a fictional or approved input for the first run.
Choose the working environment
Use Chat for exploration, Work for a deliverable and Codex for a developer workflow. Confirm local versus cloud access before attaching sources.
Give a concrete brief
Example: “Use the attached Cedar Metrics call notes to make a one-page meeting brief. Include goals, supported facts, unknowns and three questions. Do not invent account history or send anything.”
Expect a draft organized around those sections. Check each fact against the notes rather than accepting the structure as proof of accuracy.
Inspect and revise
Identify one concrete issue, such as a missing source or overly broad recommendation. Request a focused revision, then keep the final file with its inputs.
Frequently asked questions
Is ChatGPT the same as a GPT model?
No. ChatGPT is a product built around models, tools and an interface. Model names identify the engines used within products or through the API.
What is ChatGPT Work?
A task-oriented experience for producing reviewable results from sources and supported tools.
Is Codex only for programmers?
It is designed around development workflows, but operators can use it for content repositories, scripts and internal tools when they can review the resulting behavior.
Do I need to buy API access to use ChatGPT?
No. Product subscriptions and API billing are separate routes. Use the API when you need software integration.
Does local work stay entirely on my laptop?
Local tools run in a local environment, but hosted model processing can involve sending content. Review the applicable data controls.
Can cloud work continue with my computer off?
Supported cloud tasks can, provided their dependencies are available in the cloud. Local tasks still depend on the local environment.
Which model should I start with?
Start with an available default and evaluate a representative task. Change models after identifying a quality, latency or cost reason.
What is the best first task?
A bounded task with a few approved sources and an output you can check, such as a one-page account brief.
Sources & further reading
- Choose Chat, Work or Codex
- Get started with Work
- Project and file environments
- Authentication and account boundaries
ChatGPT and Codex change quickly. This page was last reviewed September 22, 2026; verify time-sensitive details against the official docs above before relying on them.