ChatGPT Work is built to finish tasks, not to chat about them. Open the Work tab in the ChatGPT desktop app instead of the Chat tab, hand it a clear goal, and it plans the steps, runs them, and returns finished files: an editable Word proposal, a two-page PDF memo, a 12-page research briefing, or a shareable web dashboard. The catch is that most of that value disappears if you use it the way you use regular ChatGPT, typing a short question, reading the reply and asking again. Below is a practical way to start, organized around five habits and one optional plugin.
From answering questions to owning tasks
Regular ChatGPT runs on back-and-forth: you ask, it answers, you ask again. ChatGPT Work is designed to take a whole task off your plate, from the first step to the final deliverable. Given a goal, it breaks the work into sub-tasks, shows its progress as it goes, and saves the output as structured files such as documents and formatted decks rather than as text in a chat window.
Because of that, small changes in how you delegate make large differences in the quality, depth and presentation of what comes back. You also do not have to trust it blindly. It shows you what it intends to do before it acts, which means you can step away instead of watching the screen or refreshing the page.
▲ From chat loops to delegated tasks
Habit 1: Give a structured goal, not a short prompt
The most common beginner mistake is a brief, vague prompt (the instruction you give the AI) such as “write a project proposal.” That leaves structure and tone open to interpretation, and the result tends to be a generic draft that you then have to edit, reorganize and reformat by hand.
A useful goal spells out four things:
| Element | What to specify |
|---|---|
| Audience | Who will read it, for example a business client |
| Structure | The sections and their order |
| Must-haves | Content that cannot be left out |
| Format | The final file type and layout |
For example, ask ChatGPT Work for an onboarding proposal for a new client that includes an executive summary, an implementation timeline, resource requirements and next steps, written for a business audience. It maps out the sub-tasks, shows each one as it runs, and produces a full business document with formatted tables and headers instead of a short paragraph. The finished proposal appears in a preview panel on the right and is saved on your computer as an editable Microsoft Word file. Choosing “Show in File Explorer” from the file’s menu takes you straight to it. Naming the destination folder and file name in your prompt makes results even easier to find.
Habit 2: Feed it every relevant file at once
Real business information rarely sits in one tidy file. The project brief is in a document, the numbers are in a spreadsheet and the latest decisions are in meeting notes. Opening each one, copying figures and reconciling notes by hand is tedious and forces constant context switching. That kind of administrative juggling is estimated to take up as much as 60 percent of a knowledge worker’s working time.
ChatGPT Work can take in several files at once, cross-reference them and merge the results into a single report. Using the plus icon in the input bar, you might attach three files together:
- A text file with the project brief
- A CSV spreadsheet of quarterly metrics
- A text file of team meeting notes
Then ask for a third-quarter performance report that compares the brief’s objectives with the metrics and folds in the action items from the meeting. ChatGPT Work reads each source, checks the percentage gaps against targets, and organizes owners and deadlines into an executive document.
Habit 3: Review the plan in Plan Mode before anything runs
A familiar frustration with complex AI tasks is discovering halfway through that the model picked the wrong format, structure or scope. Plan Mode prevents that by laying out the full execution strategy before any deliverable is generated. The plan lists the steps, the inputs each depends on, the visual format and where the files will be saved.
Suppose you use Plan Mode to request an executive competitor analysis of collaboration suites, including Microsoft 365, Google Workspace, Slack, Notion, Asana, Monday.com, ClickUp and Atlassian. ChatGPT Work first asks clarifying questions about depth and enterprise scope, then proposes a 13-slide PowerPoint deck built around a weighted five-point scorecard:
| Scorecard criterion | Weight |
|---|---|
| Security, compliance and administration | 20% |
| Productivity and collaboration breadth | 20% |
| AI capability and governance | 15% |
| Integration and extensibility | 15% |
| Work management | 10% |
| User experience | 10% |
| Commercial efficiency and consolidation potential | 10% |
If you would rather have a short document than slides, this is the moment to say so. You can edit the plan directly: change the deliverable to a strict two-page PDF briefing memo with a fixed file name, drop Atlassian from the sections on developer and ticketing workflows to keep the focus on general collaboration tools, and delete a meeting agenda section a concise memo does not need. Clicking “Yes, implement this plan” then starts the build. The resulting two-page PDF includes an executive conclusion, the weighted scorecard, three portfolio architecture patterns and a 90-day agenda. A few minutes spent on the plan catches formatting errors, extra sections and the wrong data sources before the token-heavy work begins. Tokens are the units of text a model processes, and they drive usage.
Habit 4: Let long research run in the background
ChatGPT Work runs long, multi-step jobs asynchronously. You can switch to another app, close the project window or leave your desk, and the web browsing, file analysis and document assembly continue.
For example, you can ask for a market research audit of enterprise AI agent platforms as of August 2026. An AI agent is software that takes a goal and carries out multiple steps on its own. The prompt can tell it to identify the top five platforms from technology news, vendor announcements and press releases; evaluate features, pricing and integrations using documentation and software review sites; gather market trends, financial filings, customer sentiment and common buyer complaints; and deliver a publication-ready PDF. While it works, you can minimize the window and get on with a spreadsheet in Microsoft Excel. A desktop notification tells you when it is done, and one click takes you to the finished file.
▲ Research running in the background
In this case the output was a 12-page executive briefing saved in the workspace outputs folder. It profiled Microsoft Copilot Studio, Salesforce Agentforce 360, Google Gemini Enterprise, AWS Bedrock AgentCore and ServiceNow AI Agent Platform, and included a side-by-side feature comparison, a review of commercial models, a risk and governance matrix, control checklists and a 90-day roadmap. The research, synthesis, cross-checking and layout all happened without anyone stepping in.
Habit 5: Keep permissions at the default
Unattended runs are only as safe and reliable as their permissions. In the desktop app, go to Settings, then General, and look under the Personal section.
| Setting | What it allows | Recommendation |
|---|---|---|
| Default permissions | Reads and edits files inside the designated workspace folder without asking each time; pauses and asks for approval when it needs anything outside that folder | Keep on |
| Full access | Modifies arbitrary files on the computer and runs network terminal commands automatically | Only when strictly required |
Full access removes the manual approval checkpoints for sensitive system operations, so it calls for extreme caution. The sensible setup appears to be Default permissions on for smooth unattended work and Full access off unless a specific workflow truly needs it.
Optional: Turn data into a web app with the Sites plugin
Large, multi-dimensional datasets are awkward in static PDFs and long Word files, where readers have to scroll endlessly. The Sites plugin lets ChatGPT Work turn raw data tables and Markdown memos into hosted, responsive web apps with dashboards, search filters, expandable details and interactive charts.
Turn on the Sites plugin from the plus menu, select the source files, such as a third-quarter performance CSV and a strategy brief in Markdown, and ask for an executive dashboard with top-level navigation, filterable KPI tables, collapsible analysis sections and summary charts. ChatGPT Work reads the data, loads interface frameworks, validates the metrics, writes the code and shows a preview inside the workspace. In the sharing dialog, switching access from “Just me” to “Anyone on the internet” produces a public link that opens in any standard browser.
The resulting dashboard had Overview, Portfolio, Insights and Actions tabs; KPI cards showing 10 completed, 5 on-track, 10 at-risk and 5 delayed initiatives; a stacked chart of status by department; a searchable list of initiatives filterable by department, region and status; and a collapsible strategic readout ending in seven recommended moves for leadership in the fourth quarter. Executives can scan the big picture while department leads filter down to their own line items. Static documents still make more sense for formal compliance archives, printable memos and fixed legal records.
A starting routine
ChatGPT Work pays off most when you treat it as a coworker you delegate to, not a chatbot you question. A reasonable way to begin:
- Open the Work tab in the desktop app and pick one document you produce regularly. Delegate it with a goal that names the audience, sections, must-haves and file format.
- If the inputs live in several files, attach them all at once with the plus menu and let it reconcile them.
- For large or format-sensitive jobs, start in Plan Mode, trim the plan and fix the format, then approve it.
- Send long research to the background and keep working until the notification arrives.
- Leave permissions at Default permissions and enable Full access only briefly for a task that requires it.
- Use the Sites plugin for data that several teams need to filter, and keep PDFs for fixed records.
Set a clear goal, hand over the source material, check the plan and then let it run. Those few steps alone are likely to cut much of the time spent reworking AI output.