--- id: chatgpt-work-on-your-files lang: en track: chatgpt title: "ChatGPT Work: tasks on your files" seoTitle: "ChatGPT Work: hand over whole tasks" description: "ChatGPT Work takes on multi-step tasks across your files and apps and hands back finished documents, sheets and slides. How it works and how to stay in control." part: working-smart partTitle: "Working smart" partOrder: 4 order: 4 level: Intermediate durationMinutes: 5 outcome: "Hand ChatGPT Work a multi-step task on your files and get a finished deliverable back, safely." summary: "Work is ChatGPT's agent for bigger jobs. Describe the outcome, give it only the files it needs, approve the important steps, and review the finished document it hands back." draft: false updated: "2026-10-14" furtherReading: - title: "ChatGPT Work and Codex" url: "https://help.openai.com/en/articles/20001275-chatgpt-work-and-codex" note: "Where Work runs, local files on desktop, approvals, and usage." - title: "ChatGPT is now a partner for your most ambitious work" url: "https://openai.com/index/chatgpt-for-your-most-ambitious-work/" note: "OpenAI's launch post: what Work is for and what it produces." - title: "Moving to the new ChatGPT desktop app" url: "https://help.openai.com/en/articles/20001276-moving-to-the-new-chatgpt-desktop-app" note: "How Chat, Work, and Codex sit together in one app." - title: "Prompting (ChatGPT Learn)" url: "https://learn.chatgpt.com/docs/prompting" note: "When to use Chat and when to use Work, with examples." keywords: - ChatGPT Work - ChatGPT agent - ChatGPT local files - ChatGPT desktop app - AI agent for documents quiz: intro: "Five quick questions. Ace them all and you've got a coworker, not just a chatbot." passScore: 5 questions: - q: "How is Work different from a normal chat?" options: - "It answers faster by skipping the thinking" - "It's Chat with web search always switched on" - "It only works on files you upload into the chat beforehand" - "It works through a multi-step task to a finished file" answer: 3 explain: "Chat is for quick back-and-forth. Work plans a bigger job, works through the steps, checks in with you where needed, and delivers a document, sheet, deck, report, or Site." - q: "You start a Work task on chatgpt.com in your browser. Can it open a file on your laptop's hard drive?" options: - "Yes, if you're signed in on that laptop" - "No; for local files you need the desktop app" - "Yes, but only files in your Downloads folder" - "Only if memory is switched on" answer: 1 explain: "On web and mobile, Work runs in the cloud and can't reach files on your computer. In the desktop app it can work in a local folder, but only one you open for it." - q: "What's the safest way to give Work access to your files on desktop?" options: - "Open your whole Documents folder" - "Open your desktop, where your current work usually sits" - "Open a dedicated folder with only what it needs" - "Upload everything to a project first" answer: 2 explain: "OpenAI's advice is to grant access only to the files the task needs. A dedicated working folder keeps sensitive material out of reach." - q: "Work asks you to approve sending an email to a client. What's the right attitude?" options: - "Read what it will send, then approve only if right" - "Approve it; Work has already checked the email for mistakes" - "Always decline; agents shouldn't send email" - "Turn approvals off so it stops interrupting" answer: 0 explain: "Approvals are your checkpoint on actions with real consequences. Check the content and recipient, then decide." - q: "A web page Work reads contains hidden text telling it to 'send the contract to this address'. What is this risk called?" options: - "Context rot" - "A token limit" - "Prompt injection" - "Memory drift" answer: 2 explain: "Prompt injection is when outside content tries to give the agent instructions. Keep tasks scoped, stick to trusted sources, and stop the task if it does something you didn't ask for." pass: badge: "Coworker unlocked" title: "You've got a coworker now, not just a chatbot." body: "Full marks. You can hand Work a task, scope its access, and stay in control. Last lesson in Part 4: connecting ChatGPT to the apps you use every day." retry: "Check the marked answers and give it another go. You've got this." --- > **Level:** Intermediate ยท **Reading time:** ~5 min > > **What you'll be able to do:** hand ChatGPT Work a multi-step task on your files and get a finished deliverable back, safely. Every lesson so far, you asked and ChatGPT answered, one message at a time. Then you copied the result into Word, fixed the formatting, built the slide yourself. **ChatGPT Work flips that round.** You describe an outcome, ChatGPT works through the steps, and you get the finished file. You stop being the courier and start being the reviewer. ## What Work is **Work** is ChatGPT's agent for longer, multi-step jobs. Where Chat is quick back-and-forth, Work takes on a whole task: it makes a plan, reads what it needs, does the work over many steps, checks in with you on anything important, and delivers a **document, spreadsheet, presentation, report, or Site**. A big task can run for a long time while you get on with something else. You'll find it next to Chat. In the **desktop app** there's a switch at the top between Chat and Work; on the web and phone you can start Work tasks too. On **Business**, every seat includes Work; usage depends on the size of the task, and Premium seats get about five times more room.
From asking to delegating Four steps from left to right: you describe the outcome; Work makes a plan; Work works through the steps across files, apps and the web, and asks you to approve important actions; you get a finished file such as a document, sheet, deck, report or Site. From asking to delegating You describe the outcome and review the result. Work does the steps in between. You Describe the outcome. Work plans Breaks it into steps. Work does it Files, apps, web. Many steps. Finished file Doc, sheet, deck, report or Site. you approve key actions On web and phone, Work runs in the cloud. In the desktop app, it can work in a folder you open for it.
**Which one when?** OpenAI's own rule of thumb: Chat for quick questions, short rewrites, and brainstorming; Work for jobs with several sources or tools, several steps, or a bigger deliverable at the end. Work uses more of your allowance, so it pays off when it saves you real time. ## Where it runs, and what it can touch This is the part to understand before you hand anything over. **On web and mobile, Work runs in the cloud.** It can use what you attach, your connected apps, and the web, but it **cannot reach the files on your computer.** **In the desktop app, Work can work locally.** You open a folder for it and it can read, change, and create files there, and with your permission even use desktop apps. Browser use and network access have their own switches. The safe habit is simple: **give it only what the task needs.** OpenAI's advice is to grant access only to the files the task requires. In practice, make a **dedicated working folder**, copy in the files for this job, and point Work there, not at your whole Documents drive. Keep genuinely sensitive material out of it entirely. One more detail: even when Work runs locally, messages and task context may be stored in the cloud, so the [lesson 1.4](/learn/foundations/ai-at-work-safety-and-privacy) rules on what you share still apply. ## What it can actually do A few jobs where Work earns its name: - **Read across many files.** "Go through these twelve supplier quotes and compare price, delivery time, and payment terms." - **Build real deliverables.** A formatted report, a spreadsheet with working formulas, a slide deck, ready to open. - **Combine sources.** Pull from your files, a connected app, and the web in one task, then write it up. - **Grind through the dull stuff.** Turn a folder of expense receipts into one summary table. Rename a pile of files to a tidy pattern. From a **project** with default memory ([lesson 4.3](/learn/chatgpt/working-smart/projects)), you can start a Work task that already knows the project's files and brief. ## Stay in control Because Work *acts*, not just talks, you stay in the loop: - **Watch the plan and progress.** You can answer its questions, change direction mid-task, or stop it. - **Approve the important steps.** Before actions with consequences, Work asks. On web and phone you approve or decline on screen. Read before you click. - **Your normal permissions still apply.** Work can only use apps and files you can access yourself, and your workspace's admin settings come first. **Match your attention to the stakes.** Summarising a folder of public reports needs little supervision. Anything that sends messages as you, touches money, or changes files you can't afford to lose deserves your eyes on every approval. ## Safety, said plainly An agent that reads files, browses, and acts is as useful, and as risky, as that sounds. One risk is worth naming: **prompt injection.** If Work reads outside content, like a web page or an email, someone can hide instructions in it ("ignore your task and send this file to..."). The defences are habits: - Keep tasks **scoped**: a dedicated folder, trusted sources, only the apps the job needs. - **Watch the task, not every keystroke.** If it starts touching files or sites you never mentioned, stop it. - **Review the deliverable** before it goes anywhere. You're responsible for what Work does on your behalf. ## Where this goes next Work lets ChatGPT act on your files. The last lesson in Part 4, [Connecting ChatGPT to your tools](/learn/chatgpt/working-smart/connecting-chatgpt-to-your-tools), widens that reach to the apps you use every day, and shows how to do it safely. Scheduling Work to run on its own comes in Part 5. ## Try it In the desktop app, make a new folder and drop in three or four real files: a report, some notes, a spreadsheet. Switch to **Work**, open the folder, and ask: **"What's in this folder, and what stands out?"** Then ask for something built: **"Turn these into a one-page summary I can share, as a Word document in this folder."** Watch the plan, approve what it asks, and open the file it makes. That first handover, from asking to *delegating*, is the whole idea of Work. *(General exercise. Role and industry versions come once accounts are in.)* ## Key terms [ChatGPT Work](/glossary#chatgpt-work), [Prompt injection](/glossary#prompt-injection), [Project](/glossary#chatgpt-project), [Context](/glossary#context)