ChatGPT Work 实战指南：构建人机协作的安全大本营

ChatGPT Work | Get real work done with ChatGPT Skip to content

ChatGPT Work GuidePractical work with ChatGPT WorkOverviewFeaturesSitesRemotePlaybookPromptsFor Managers

Agent mode

Menu

Overview↗Features↗Sites↗Remote↗Playbook↗Prompts↗For Managers↗Agent mode

One home base for all your work

A practical guide to using ChatGPT Work on desktop.

Set up your home baseExplore features

Finance example

SlackThe request

ExcelCurrent source

Google DocsHistorical context

Sites

Forecast Review Site

Draft for your review

FinanceLegalSalesDesignDataEducationMarketingPeopleProductStrategy Ops

Your roleFinanceLegalSalesDesignDataEducationMarketingPeopleProductStrategy Ops

Review a forecast interactively

Compare the supplied retention cases and test changes to forecast months. The source workbook stays unchanged.

Sites · Example

Forecast Review Site

ForecastSlidesSources

Current outlook

$85.8M

FY26 total revenue +$200K vs plan

Drag any forecast month · up to $5M

Monthly revenue: actuals through June and the selected forecastActualsForecastMonthly revenue ($M)$20M$15M$10M$5M$0M$14.2MJul: $7.5M monthly revenueAug: $8.7M monthly revenueSep: $10.0M monthly revenueOct: $11.3M monthly revenueNov: $12.7M monthly revenueDec: $14.2M monthly revenueJanFebMarAprMayJunJulAugSepOctNovDec

Monthly revenue: actuals through June and the selected forecastActualsForecastMonthly ($M)$20M$15M$10M$5M$0M$14.2MJul: $7.5M monthly revenueAug: $8.7M monthly revenueSep: $10.0M monthly revenueOct: $11.3M monthly revenueNov: $12.7M monthly revenueDec: $14.2M monthly revenueJanMarJunSepDec

Week-4 retention scenario

40% lower50% current60% higher

Reset to current outlook

Decision: keep the original outlook until the cohort evidence and rollout assumptions are reviewed.

Executive summary

The $9.9M higher case needs experiment evidence

The $85.8M current outlook remains the planning case pending review.

FY26 revenue outlook

$85.8MAnnual plan $85.6M

Current caseJan–Jun actuals are $21.3M. The Jul–Dec forecast contributes $64.5M.

Higher caseThe supplied 60% retention case totals $95.7M, or $9.9M above the current outlook.

DecisionReview mature cohort evidence and rollout assumptions before adopting the higher case.

Synthetic example · Original outlook Sources: Slack, Excel, Google Docs

Fictional sources for this example

Slack#business-review

ExcelFY26 revenue outlook.xlsx

Google DocsLast quarter’s business memo

FY26 revenue outlook.xlsx

FY26 revenue outlook.xlsx · Supplied exampleFY26 plan$85.59MFY26 current outlook$85.79MJan–Jun actuals$21.30MJul–Dec forecast$64.49MCurrent retention case50%

Fictional first-year rollout of a subscription product, in USD millions. Jan–Jun actuals total $21.30M; Jul–Dec forecast totals $64.49M. FY26 outlook is $85.79M versus an $85.59M plan. The steep ramp assumes expanding product availability; those assumptions are not validated here. The supplied 40%, 50%, and 60% retention cases total $78.82M, $85.79M, and $95.70M. These are model cases, not observed outcomes or a causal formula. Monthly values between cases are interpolated. Dragging a forecast month adds a separate adjustment of up to $5M, never below zero; only that month and the annual sum change. Sources and review slides stay unchanged.

How the draft uses itProvides the monthly inputs, exact annual control, and three cases. Manual adjustments explore a separate assumption; they do not revise the workbook.

Try this with your work

Use approved app connections or add the files directly.

These walkthroughs use the desktop app. Menus and available features may differ in the browser.

• 1

Bring your sources

Bring the current forecast, the Product question, and the previous review.

• 2

Ask for a first draft

Open the prompt in ChatGPT desktop or copy it into a task with your sources. Name the audience and result you need; add a template or past example if you have one.

• 3

Review what matters

Check the cohort definition, observation window, and rollout assumptions before treating the higher case as the forecast.

Copy Finance promptOpen in ChatGPT desktop ↗

Official OpenAI walkthrough | Finance

Explore a forecast interactively

Forecast and product context → interactive scenario review. Check proposed changes against the original model before approving an update.

Watch official walkthrough

Duration 1:37

Watch on YouTube ↗

AGENTS.LL.md|realwdsourhumanreviy→revixamples|retereceiptsforeactionmproveoChatGPTWork|appr.md|realwtedsouowner|humanrrce→verireshexampltaff|clpprovalbefoow|imChatGPTWoAGENTS|realworkfirrces|verowner|humaurce→veexamples|re|cleaeforeacti

Learn the method. Bring your agent up to speed.

Paste into a new ChatGPT Work chat

I’m just getting startedOpen a desktop draft ↗

Map your work and design a useful setup.

Evaluate my current setupOpen a desktop draft ↗

Review all six stages and prioritize improvements.

Copy buttons include the prompt and complete guide. Paste into a new chat. Desktop links prefill a draft with the guide link, not an attachment; you still send it. Nothing is inspected or connected here. Download the full guide (.md)

1 | Mindset

Why make it your home base?

Your work can build on what you have already taught it. Bring context together, review the result, and keep the methods that make the next task better.

Context you can return to

Keep source files, project instructions, and useful notes together. Where memory is available, choose which preferences to retain.

Corrections that carry forward

Review a real draft, explain what should change, and save the useful corrections in the project. The next task starts with better guidance.

Skills you have taught and tested

Capture a method that worked, test it on fresh examples, and reuse it. Each review is a chance to improve the instructions.

Work you can plan ahead

Agree on future checks or recurring drafts. With approved access and a schedule, it can look for relevant updates and prepare the next useful step.

• Do real work

• Review and correct

• Save what works

• Test and reuse

Save the corrections you want the next task to use. Recheck the sources and test the method on new work; keep important decisions with you.

2 | Context

Give it the context to do the work

Start with one useful project and the files and notes you’re allowed to share.

• 1

Choose a job you already know

A review, a memo, or a team update. Name the result and who needs it.

• 2

Connect the context that matters

Use approved files, Slack conversations, and meeting notes. Check what it can read, what’s current, and what’s missing.

• 3

Give the work a home

Keep sources, instructions, and corrections in the existing project. Reuse a tested Skill when it fits.

ChatGPT Work Plugins

Start with the right tools

• DocumentsDraft a memo from your source notes

• SpreadsheetsBuild a workbook and check formulas

• PresentationsTurn your brief into a presentation

• SitesBuild an interactive dashboard, slides and more

In the app, open Plugins → OpenAI.

Select an installed plugin with @ in a new chat.

Availability depends on your workspace.

Check my work setup ↗Take the same work with you on your phone →

3 | Workflow

Define the task, sources, and checks

Specify the result, where the information lives, and what you need to review.

• 1

Trigger

What starts this work?

• 2

Trusted sources

Where does the right information live?

• 3

Expected output

What should the finished result look like?

• 4

Quality checks

How will you know the work is right?

• 5

Human decision

What still needs your approval?

Copy the first-task prompt →

4 | Skill

Turn repeat work into a Skill

A Skill is a reusable set of instructions for work you do repeatedly. You teach it the process, sources, and checks through a real example and your corrections, then test it on fresh work.

Draft while the real work is fresh. Use the example and your corrections to capture the sources, steps, and checks, then test the Skill with new inputs.

• 1

Choose recurring work

Start with a real job you expect to do again.

• 2

Share the context

Show the real sources, audience, and useful examples.

• 3

Work through a real example

Ask ChatGPT to help with the actual work.

• 4

Correct what matters

Explain what should change and which decisions stay with you.

• 5

DRAFT THE SKILL HERE

Capture the method while the real work is still in context.

• 6

Test on fresh examples

Check that the draft works with new information.

• 7

Improve and reuse

Keep what holds up and fix what does not.

A Skill saves the method. It does not connect tools, start a schedule, or give permission to act. Set those up separately when you need them.

5 | Automation

Plan the next useful check

Once the method holds up, agree on when it should run. An approved schedule can turn new information into a draft or a short list of decisions for you.

• Name the signal and timing

For example, review approved team updates each morning or prepare a draft before a recurring meeting.

• Define the allowed work

Specify the sources, the draft you want, the checks to run, and where to keep the result.

• Set the stop conditions

Pause for missing access, conflicting evidence, or a decision that needs you. Sending, publishing, and changing sources need separate approval.

Connecting an app alone does not start monitoring. Set up the schedule and permissions explicitly, then review whether each run is useful.

See a practical daily workflow →

6 | Operating system

Review your projects from one home base

Keep a short list of active projects, their owners, decisions, and next steps. Return to each project for its sources and detailed work.

One project per workstreamKeep the owner, source files, instructions, and ongoing work together.→Instructions that fit the jobKeep your general preferences separate from the instructions for this project.→App Shots with your permissionUse App Shots or Computer History only when available, relevant, and approved.→Skills that stay usefulCapture personal or approved shared methods, test fresh examples, and review them monthly.→Status checks that show what changedCheck what actually advanced without confusing a status check with approval.→Review progress across projectsPrepare a brief from approved updates, with the owner and next decision for each project.→

Keep these things true: read the real source, automate what you understand, and keep the final decision. A draft is not permission to send, publish, deploy, or change access.

Review the six stages →Find a chapter →Find a useful prompt →Build a Site →Find your starting point

Choose the route that fits your work. Your selection stays on this device.

Run my workProjects, sources, tasks, checkpoints, and a personal Chief of Staff.Build the systemTurn workflows into Skills, validators, and bounded automations.Scale my teamOwners, approval gates, collaboration, and recovery across a team.

Choose a path to reveal a four-step reading sequence.

Try it with your workStart a five-day team pilot

Official walkthroughs

How-to videos

Pick a short walkthrough from OpenAI’s teams. The interactive examples above use fictional content.

Watch the full playlist ↗ - Duration 1:48StrategyBuild a strategy deckWatch on YouTube ↗

• Duration 2:33MarketingGet a solid blog draftWatch on YouTube ↗

• Duration 1:41SalesPrepare for customer meetingsWatch on YouTube ↗

• Duration 1:34Data ScienceSchedule a weekly metrics reportWatch on YouTube ↗

• Duration 1:37FinanceExplore a forecast interactivelyWatch on YouTube ↗

Templates30-day planChief of StaffDownloadsAgent instructions (.md)About ChatGPT Work ↗Reduce motion