Getting started with ChatGPT

ChatGPT offers different ways of working. Choose the environment that suits your task:

Use Chat when you mainly want to discuss something. Switch to Work when ChatGPT needs multiple steps, sources, or tools to deliver a result. Choose Codex for software development or when you want to follow technical details, files, and changes precisely.

Which parts you see depends on your subscription, device, region, rollout, and any settings applied by your organization.

You do not need a magic prompt. A useful instruction usually contains four parts:

• Goal: what should ChatGPT achieve?
• Context: which background information, files, links, or examples are relevant?
• Output: in what format, length, and tone do you want the result?
• Boundaries: what must ChatGPT not do, and what should you approve first?

Use this starting prompt and replace the text in brackets:

Goal:
[Describe the concrete result you need.]

Context:
[Provide the relevant background, sources, files, and target audience.]

Desired output:
- Format: [for example, table, email, report, presentation, or website]
- Length: [desired scope]
- Tone: [desired writing style]
- Required sections: [list]

Boundaries and review:
- Do not invent facts, names, figures, links, or quotes.
- Point out missing or uncertain information.
- Ask for my permission before sending, publishing, purchasing, changing, or deleting anything.
- Check the final result against [criteria].

Your first message does not need to be perfect. Ask ChatGPT to identify missing information, have it create a first version, and then give specific feedback. “Make the opening shorter and start with the conclusion” works better than “This is not good enough yet.”

Upload documents, spreadsheets, PDFs, images, or data when ChatGPT genuinely needs them. For example, ask it to:

• compare two quotations in a table;
• turn meeting notes into decisions and action items;
• find patterns or errors in a spreadsheet;
• summarize a long document for a specific audience;
• combine information from multiple files into a new report.

Create a Project when several conversations need the same files, sources, and instructions. Start a separate chat within that project for every clear deliverable. This keeps the context organized without requiring you to upload the same background information repeatedly.

Pin projects and chats you return to often. Give them a short, recognizable name and archive completed work. You can later use search to find words from a conversation, even when you no longer remember its title.

Use web search or the browser when current information, multiple websites, or interactive pages are part of the task. You can ask ChatGPT to compare products, collect prices and terms, and put the results in a spreadsheet.

A good browser instruction specifies:

• which question needs to be answered;
• which websites or types of sources may be used;
• which information needs to be collected;
• how the result should be presented;
• which actions require your permission first.

In Work on the web, ChatGPT can use a separate cloud browser. In the desktop app, you can use the built-in browser or, where available, let it control a Chrome tab. The built-in browser has its own profile; a Chrome connection can work with pages where you are already signed in.

A website may contain misleading instructions. Always check the domain name and the proposed action. Keep control of purchases, forms, account settings, messages, and data deletion.

In supported situations, Computer Use allows ChatGPT to view windows, click, type, and operate apps. This is useful for a clearly defined action, such as organizing files, transferring information from an application, or carrying out a routine check.

Start with one application and one concrete goal. Keep sensitive apps closed, do not share passwords in a prompt, and stay present when working with privacy, security, payment, and account settings. Review the result before allowing files to be deleted, data to be sent, or settings to be changed.

In the ChatGPT desktop app, you can start a new conversation directly in Voice. This lets you discuss ideas, start a task, ask for progress updates, and adjust the direction without typing. Voice can also start separate tasks and bring their results back into the voice conversation.

To do this, open a new, empty chat and choose Start new voice chat. Grant microphone permission the first time you use it. On macOS, you can optionally enable screen context, but remember that a screen recording may also contain sensitive text outside the currently visible part of a window.

Voice uses the same permissions and boundaries as ordinary tasks. Talking to ChatGPT does not automatically give it more access.

In Work, ask not just for text but for the file you ultimately need: a document, presentation, spreadsheet, PDF, or interactive visualization. Open the result and review every tab, page, or slide before downloading or sharing it.

Sites lets you create and host a website, dashboard, web app, or simple game. Describe at least:

• who the site is for;
• what a visitor should be able to do;
• which content and files should be used;
• which information needs to be retained;
• who should be allowed access.

A new Site initially has restricted access. Review its text, links, forms, uploads, personal data, and any secret values before widening access. Publish only after testing the site as a visitor.

Start with the default Power setting when you are not sure which model you need. If you want to choose for yourself, use this practical breakdown:

• GPT-5.6 Sol: difficult, open-ended tasks that require substantial analysis, judgment, or refinement;
• GPT-5.6 Terra: everyday work with a good balance of quality, speed, and usage;
• GPT-5.6 Luna: clear, repeatable tasks such as organizing, classifying, converting, and creating standardized summaries.

Start with the standard reasoning effort. Increase it only when the task requires more planning, sources, trade-offs, or review. Max gives one model extra room for an exceptionally difficult problem. Ultra can divide parts of the work among multiple agents in parallel. Most everyday tasks do not need Max or Ultra.

A more powerful model is not automatically better for every step. For example, you can have Sol design a complex approach and use Luna for small, clear changes or repetitive work.

Build automation in this order:

1. First perform the task manually a few times.
2. Incorporate recurring corrections into a standard workflow.
3. Create a Skill when you need the same approach more often.
4. Only then create a Scheduled task if the work also needs to run at fixed times.

A Skill contains instructions, examples, and optionally tools for a recurring process. A plugin is an installable package that can combine skills with connections to services such as Google Drive, Gmail, GitHub, Notion, or Slack.

Connect only a service that the task requires. Check which data the connection is allowed to read and which actions it can perform. Installing a plugin does not change the access permissions of the connected service: ChatGPT can only work within the permissions of the account you use to connect.

Under Scheduled, you can create recurring tasks and review completed runs. Suitable first examples include:

• an overview of your calendar, unread messages, and priorities every workday;
• a weekly report based on fixed sources;
• a daily check that a website is available;
• periodic summaries of log files or project updates;
• collecting maintenance suggestions at a fixed time.

Describe what each run should check, which sources it may use, when it should report something, and when it should stop or ask for help. Test the prompt in an ordinary chat first and review the initial runs carefully.

A local scheduled task requires the computer to be switched on and the desktop app to be running. Web tasks can work with uploaded information and connected services, but not directly with an arbitrary folder on your computer. For unattended tasks, use the narrowest access permissions that still allow the task to succeed.

Choose Goal mode when ChatGPT needs to work through multiple steps independently until it reaches a verifiable final result. In the desktop app or Codex, type /goal and describe:

• the desired result;
• the boundaries and prohibited actions;
• how ChatGPT can prove that the work is complete;
• a time, cost, or usage limit if that is important.

You can pause, resume, or redirect an active goal. Goal mode does not give ChatGPT broader permissions; existing sandbox, network, and approval rules remain in effect.

Remote lets you monitor and control work on a connected Mac or Windows computer from the ChatGPT mobile app. You can open existing chats, send new instructions, answer questions, grant approvals, and view results or changes.

Start the connection in the desktop app through Set up Remote or Settings > Connections, then scan the QR code with your phone. Both devices must be signed in to the same ChatGPT account and workspace. The host computer must remain awake, online, and signed in.

Remote uses the files, apps, browser, connections, permissions, and security settings of the connected computer. Treat an approval on your phone exactly as if you were sitting at that computer.

Models, Voice, Sites, and agent tasks can have different usage limits. Check your current usage through your profile or the usage dashboard. If you are using your budget faster than expected, choose a smaller model, limit the scope of the task, or split a broad assignment into clearer steps.

Never let cost be the only quality check. Important tasks still require source verification, human review, and a clear definition of “done.”

1. Open ChatGPT on the web or in the desktop app.
2. Choose Chat, Work, or Codex based on the result you want.
3. Select one task whose result you can evaluate yourself.
4. Write your instruction with a goal, context, output, and boundaries.
5. Add only the files, websites, or connections that are needed.
6. Review the result and give one round of specific feedback.
7. Turn it into a Skill or scheduled task only when the approach has proved reliable.