How to Use ChatGPT Work: The Complete Guide

ChatGPT Work turns a clear goal, your files, connected tools, and approved actions into finished documents, spreadsheets, presentations, analyses, workflows, and Sites. This complete guide explains how to use it properly.

How to Use ChatGPT Work: The Complete Guide

Updated July 12, 2026: ChatGPT Work is still rolling out across plans and surfaces. Availability, menu labels, model controls, plugins, and workspace permissions may differ slightly depending on your account.

ChatGPT just grew a second gear.

Regular Chat is where you ask ChatGPT a question, work through an idea, brainstorm, search the web, or knock out a quick draft.

ChatGPT Work is where you hand it a job.

Not one tiny prompt.

Not one disposable answer.

A job with source material, multiple steps, tools, decisions, files, review criteria, and an actual finished result at the end.

OpenAI describes Work as an agent for substantial tasks. It can gather information from your files and connected tools, break a larger goal into smaller pieces, work across apps and websites, create editable deliverables, run recurring workflows, and stay with a project for hours when needed.

That’s a much bigger shift than adding another button to the ChatGPT interface.

For years, most people used ChatGPT like a brilliant person trapped behind a text box. You could ask it to explain something, draft something, or analyze whatever you pasted into the conversation.

Work gives it a desk.

Files on one side. Connected apps on the other. A browser, models, permissions, reusable workflows, scheduled tasks, and a growing list of formats it can produce.

The basic concept is easy:

Chat gives you a response. Work gives you a result you can review and use.

The hard part is learning how to delegate well enough that the result is actually useful.

This guide covers the entire system, including how to find Work, choose the right mode, structure a task, provide files and constraints, connect plugins, create documents and spreadsheets, use the desktop browser, schedule recurring work, build Sites, choose a GPT-5.6 model, manage credits, and keep control of important actions.

What Is ChatGPT Work?

OpenAI defines ChatGPT Work as an agent inside ChatGPT that can take action across your apps and files, stay with a project for hours if needed, and turn a goal into finished work.

That last phrase matters.

Work is designed around outcomes.

You aren’t only asking it to tell you how to create a competitor analysis. You can ask it to gather the approved source material, compare the competitors, build the spreadsheet, create the executive summary, generate a presentation, and stop with everything ready for review.

A Work task can include:

  • Research
  • Uploaded files
  • Local files on desktop
  • Connected apps
  • Plugins and Skills
  • Web browsing
  • Computer Use
  • Documents
  • Spreadsheets
  • Presentations
  • Reports
  • Analyses
  • Scheduled workflows
  • Interactive Sites
  • Human approval before sensitive actions

You can follow the task while it runs, answer questions, redirect it, add information, and approve or deny important actions.

Work is powered by Codex technology, but it isn’t simply Codex with a new label.

Codex remains the dedicated environment for software development. Work takes the same general agentic machinery and points it at everyday knowledge work.

Chat vs. Work vs. Codex

The new ChatGPT desktop app places three different working styles beside each other.

ModeBest forTypical result
ChatQuestions, explanations, brainstorming, web search, comparisons, and shorter draftsAn answer or conversational draft
WorkResearch, analysis, multi-step tasks, files, workflows, and polished deliverablesA document, spreadsheet, deck, report, Site, or completed workflow
CodexSoftware development using repositories, local folders, terminals, and developer toolsWorking code, tested changes, reviews, or a shipped technical task

The dividing line isn’t the number of words in your prompt.

It’s the type of responsibility you’re handing over.

Use Chat when you want to think with ChatGPT.

Use Work when you want ChatGPT to complete a defined piece of work.

Use Codex when the center of gravity is a codebase, repository, terminal, build system, or software deployment.

Here is a practical example.

Use Chat

What should a good monthly marketing report include?

Use Work

Use the attached campaign exports, our connected analytics data, and last month's report template to create the June marketing report.

Preserve the template structure. Explain the five largest changes, identify any data gaps, create the required charts, and stop with an editable document and spreadsheet ready for my review.

Use Codex

Add the missing campaign attribution report to this repository, write tests for the calculation logic, run the test suite, and show me the final diff.

The first asks for knowledge.

The second delegates a deliverable.

The third delegates software work.

Who Can Use ChatGPT Work?

At launch, OpenAI says the new ChatGPT desktop app includes Chat, Work, and Codex on Windows and macOS across every plan, including Free.

Work on the web and mobile is rolling out to eligible paid plans. OpenAI began with Pro, Enterprise, and Edu, followed by Plus and Business.

Rollouts are rarely clean little light switches.

Your access can still depend on:

  • Your plan
  • Your region
  • Whether the rollout has reached your account
  • Whether you’re using desktop, web, or mobile
  • Workspace settings
  • Your role in a managed workspace
  • Whether an administrator has enabled the required plugins or apps

OpenAI’s desktop migration guide also explains that the former Codex desktop app becomes the new ChatGPT desktop app after updating.

If you used the older ChatGPT desktop app, you may temporarily see two apps:

  • ChatGPT: The new app with Chat, Work, and Codex
  • ChatGPT Classic: The previous desktop app

Existing Codex projects should remain after the update.

Desktop Work and Cloud Work Are Different

This is one of the most important launch details.

ChatGPT Work on web and mobile runs in the cloud.

ChatGPT Work in the desktop app can also use local files, local folders, desktop apps, and the built-in browser when you grant permission.

Those two environments aren’t fully synchronized at launch.

According to OpenAI’s Work and Codex help page:

  • Work conversations created on web or mobile stay in the cloud.
  • Those cloud Work conversations don’t currently appear inside desktop Work.
  • Desktop Work threads stay on that computer.
  • Local files and local outputs remain on that computer unless you explicitly move or share them.
  • Regular Chat conversations continue syncing normally.

That means you should choose the surface before you begin a large task.

Use cloud Work when you need access from web or mobile, want the task to continue remotely, or don’t need local computer files.

Use desktop Work when the task depends on folders, desktop applications, local documents, local development pages, or the built-in browser.

Don’t start a giant cloud task assuming you can open the same Work thread in the desktop app later.

Right now, you can’t.

How to Start Your First ChatGPT Work Task

On desktop:

  • Open the new ChatGPT desktop app.
  • Sign in.
  • Start a standalone task, open a Project, or open a local folder.
  • Select Work from the mode switcher.
  • Describe the outcome you want.
  • Add the files, sources, constraints, and review criteria.
  • Review the result and request changes in the same thread.

On web or mobile:

  • Open ChatGPT.
  • Select Work from the mode selector.
  • Start a new conversation or open a Project.
  • Attach the source material or connect the approved tools.
  • Describe the finished deliverable.
  • Follow the progress and answer questions when needed.
  • Review the output before using or sharing it.

OpenAI recommends starting with a job you already understand.

That’s smart advice.

Your first task shouldn’t be “run my entire company.”

Pick something you already do often enough to recognize whether the result is good.

Examples include:

  • Turn meeting notes into a polished project brief.
  • Compare five products in a spreadsheet.
  • Convert research into a presentation.
  • Rebuild a recurring report from a known template.
  • Gather updates from connected tools and prepare a weekly summary.
  • Analyze a month of operating data and explain the largest changes.

The goal isn’t to see whether Work can produce something impressive.

The goal is to learn how it handles a workflow you can judge.

The Five-Part Formula for a Strong Work Prompt

A normal ChatGPT prompt can survive a little fuzziness.

A delegated job needs a contract.

The strongest Work prompts define five things:

  • Outcome
  • Sources
  • Constraints
  • Quality criteria
  • Approval boundary

1. Outcome

Describe what should exist when the task is finished.

Weak:

Look into our competitors.

Better:

Create a comparison spreadsheet and a two-page executive brief covering the five competitors listed in the attached document.

The word “research” describes an activity.

A spreadsheet and brief describe deliverables.

2. Sources

Tell Work where the truth should come from.

Examples:

  • Attached files
  • A Project
  • A local folder
  • Google Drive
  • Slack
  • SharePoint
  • Gmail
  • A list of approved websites
  • A connected CRM
  • The built-in browser

When source boundaries matter, say so plainly.

Use only the attached files and the official product pages linked in the source document. Do not use customer reviews, social media posts, or third-party summaries.

3. Constraints

Define what it must preserve, avoid, or leave unchanged.

Examples:

  • Keep all spreadsheet formulas.
  • Preserve the existing slide master.
  • Don’t change the worksheet names.
  • Use the supplied brand colors.
  • Don’t invent missing numbers.
  • Separate confirmed facts from assumptions.
  • Keep the report under 1,500 words.
  • Don’t contact anyone.

Constraints keep a polished result from becoming the wrong result.

4. Quality criteria

Explain what good looks like.

The final brief should be readable by an executive in under five minutes, cite every external claim, identify conflicting data, and end with the three decisions leadership needs to make.

You can also specify:

  • Required sections
  • Columns
  • Charts
  • Formulas
  • Slide types
  • Tone
  • Audience
  • Citation style
  • Verification checks
  • Acceptance criteria

5. Approval boundary

Tell Work which actions require you.

You may research, create files, and prepare drafts. Stop before sending messages, changing permissions, publishing anything, purchasing anything, or modifying an external record.

That last sentence can save you from a very exciting afternoon.

A Reusable ChatGPT Work Prompt Template

Complete the following task in ChatGPT Work.

OUTCOME
Create:
- [Deliverable 1]
- [Deliverable 2]
- [Deliverable 3]

PURPOSE
These deliverables will be used for:
[Audience, decision, meeting, publication, or workflow.]

SOURCES
Use:
- [Attached files]
- [Connected apps or plugins]
- [Approved websites]
- [Local folders or templates]

Do not use:
- [Prohibited sources]
- [Unverified claims]
- [Outdated versions]

PROCESS
1. Review all supplied material before creating anything.
2. Extract the confirmed facts, requirements, and open questions.
3. Build a brief plan and begin the work.
4. Flag conflicts or missing information instead of inventing answers.
5. Create the requested deliverables.
6. Verify the result against the acceptance criteria.

CONSTRAINTS
- Preserve: [layout, formulas, branding, structure, filenames]
- Do not change: [protected elements]
- Do not invent: [numbers, quotes, dates, claims, experience]
- Keep the output: [length, format, tone]

ACCEPTANCE CRITERIA
The work is complete only when:
- [Criterion 1]
- [Criterion 2]
- [Criterion 3]
- Every factual claim is supported by an approved source.
- Every required file opens correctly.
- Any limitation or unresolved question is clearly listed.

APPROVAL BOUNDARY
You may research, analyze, and create drafts or files.

Stop and ask before:
- Sending or publishing anything
- Making a purchase
- Deleting data
- Changing permissions
- Submitting information
- Modifying an external system

You don’t need this much scaffolding for every job.

For high-value tasks, it’s worth it.

Using Work With Projects

A Project keeps related conversations, files, and instructions together.

That makes Projects a natural home for ongoing Work tasks.

A client launch Project could contain:

  • The contract
  • Brand guidelines
  • Meeting notes
  • Product information
  • Previous drafts
  • Research
  • Approved messaging
  • A standing instruction file

Then each Work task can reuse that context without rebuilding the entire pile from zero.

Projects are especially useful when:

  • The work spans multiple sessions
  • Several deliverables share the same source material
  • You need persistent project instructions
  • You want related files organized together
  • You’re iterating on a report, campaign, plan, or analysis

One limitation is easy to miss:

OpenAI’s Scheduled Tasks documentation says a task created inside a Project can’t currently access that Project’s files when the scheduled run happens.

So Projects and Scheduled Tasks can work beside each other, but don’t assume a scheduled run automatically inherits every uploaded Project file.

Put the required context directly into the task or use connected sources the task can reach.

Using Plugins, Apps, and Skills

On July 9, 2026, OpenAI replaced the App Directory with the Plugin Directory.

The vocabulary is easy to scramble, so here is the clean version.

App

An app connects ChatGPT to outside data or actions.

An app might let ChatGPT search Google Drive, read Slack, inspect a CRM, use Gmail, update a record, or display information from another service.

Skill

A Skill contains reusable instructions for how to complete a task.

It can define the process, required inputs, output format, guardrails, and verification steps.

Plugin

A Plugin packages capabilities for a workflow.

It can contain:

  • Skills
  • Apps
  • App templates
  • Supporting workflow components

A Plugin can be broad, or it can be built for one line of work such as sales, finance, operations, marketing, or data analysis.

The app provides access.

The Skill provides the method.

The Plugin bundles the useful pieces together.

Open the Plugin Directory and inspect the listing before installing anything. Check:

  • What it can do
  • Which apps it includes
  • Which apps are required
  • What setup is needed
  • Which accounts it will connect
  • What actions it can take
  • Whether your administrator has enabled it

You can often call a connected tool or Plugin directly by typing @ and selecting it in the composer.

For example:

@Google Drive

Find the approved product messaging, the latest launch plan, and the customer research summary. Use those files to create a campaign brief.

Exact availability depends on the Plugin, plan, region, surface, workspace settings, and your role.

Creating Documents, Spreadsheets, and Presentations

ChatGPT Work can create or edit documents, spreadsheets, presentations, reports, and analyses.

Here’s an example — the first document I created with ChatGPT Work:

You can begin with:

  • Written instructions
  • Uploaded source files
  • An existing document
  • A master deck
  • A reusable template
  • A connected Google Workspace file

A good file request should identify:

  • The output format
  • How the file will be used
  • The source material
  • The structure
  • What must remain unchanged
  • What needs to be checked before completion

For a spreadsheet, specify:

  • Sheet names
  • Required columns
  • Formulas
  • Charts
  • Filters
  • Data validation
  • Assumptions
  • Formatting rules

For a presentation, specify:

  • Audience
  • Story
  • Sections
  • Number of slides
  • Required slide types
  • Brand template
  • Visual checks
  • Sources
  • Speaker notes

For a document, specify:

  • Audience
  • Length
  • Required headings
  • Tone
  • Citation style
  • Tables or callouts
  • Decision or action the document should support

Reference file versus reusable template

OpenAI distinguishes between a reference file and a template.

A reference file is an example for the current task.

Use this deck's master slides, visual style, and section order, but replace the content with the attached research.

A template combines an example file with workflow instructions for repeated use.

Use a reference when you need one deliverable to match one example.

Use a template when you need the same type of deliverable repeatedly.

Google Workspace

When the relevant Google Workspace apps are enabled, Work can create or edit native Google Docs, Sheets, and Slides.

You choose the account or file and approve the action before the change is made.

Microsoft files

There’re a few launch wrinkles.

Work can create or edit spreadsheet files, but direct interaction with an open Excel workbook uses Codex in the desktop app with the ChatGPT for Excel add-in.

OpenAI also says PowerPoint isn’t included in the Work desktop file flow at launch, although a separate ChatGPT for PowerPoint experience exists.

The product family is converging, but not every file workflow lives behind the same button yet.

Working With Local Files on Desktop

Desktop Work can use local folders and files after you grant access.

That’s the point where Work stops behaving like a website and starts behaving more like a local operator.

A useful local task might look like this:

Use the files in this folder to create a project handoff package.

Read every Markdown file, spreadsheet, and PDF. Create:
1. A project summary
2. A decision log
3. A list of unresolved issues
4. A table of owners and deadlines
5. A clean handoff document

Do not modify the source files. Write all outputs to a new folder named handoff-review. Stop before deleting, moving, or renaming anything.

Grant the narrowest access the task needs.

Don’t hand it your entire home directory because one spreadsheet lives somewhere inside it.

Local files and outputs remain on the computer unless you explicitly move or share them.

Using the Built-In Browser

The desktop app includes a built-in browser for Work and Codex.

You and ChatGPT can see the same page. ChatGPT can move between tabs, inspect pages, download files, wait for you to sign in, and use Computer Use to click, type, take screenshots, and verify results.

Open it from the toolbar or use:

  • Windows: Ctrl+Shift+B
  • macOS: Command+Shift+B

The built-in browser uses its own browser state.

It doesn’t automatically use your normal Chrome profile, cookies, open tabs, or logged-in sessions.

That separation is useful for control, but it also means you may need to sign in again.

OpenAI recommends entering credentials inside the browser, never in the chat.

The browser can be useful for:

  • Gathering information from approved websites
  • Reviewing a web dashboard
  • Comparing products or services
  • Downloading reports
  • Testing a local website
  • Inspecting a web app
  • Following a multi-page workflow
  • Checking the final result on the page

ChatGPT asks before using a new website unless you already allowed it or selected more permissive settings.

It also asks for confirmation before sensitive actions such as:

  • Submitting information
  • Making a purchase
  • Changing permissions
  • Deleting data

At launch, Computer Use can’t automate file uploads through the built-in browser.

Built-in browser versus cloud browser

The built-in browser runs inside the desktop app. You can watch, annotate, and steer the page directly.

The cloud browser runs remotely and can continue supported web work in the cloud.

Use the built-in browser when local visibility and control matter.

Use cloud Work when you need a delegated web task to continue away from the desktop app.

Using Scheduled Tasks

Scheduled Tasks turn Work from a one-time agent into a recurring operator.

A task can:

  • Run once at a future time
  • Repeat on a schedule
  • Check for a condition
  • Monitor something for changes
  • Notify you only when there’s something meaningful to report

OpenAI gives examples such as:

  • Reviewing new Slack updates each week and refreshing a meeting agenda
  • Checking websites and dashboards each morning
  • Monitoring customer feedback for recurring themes
  • Updating a presentation when new email feedback arrives

A strong scheduled prompt must be self-contained.

The agent may run later without the conversational context that made the request obvious to you.

Weak:

Check that thing every morning.

Better:

Every weekday morning, check the official OpenAI release notes and Help Center for new ChatGPT Work changes published since the previous run.

Report only meaningful changes to availability, models, pricing, plugins, Scheduled Tasks, Sites, desktop behavior, browser capabilities, or file support.

For each change, include:
- What changed
- The official source
- Who it affects
- What action I should consider

If there are no meaningful changes, do not notify me.

Manage scheduled tasks from the Scheduled page on ChatGPT web or mobile.

At the time of writing, OpenAI says the Scheduled tab isn’t available inside the desktop app.

Active task limits vary by plan.

Monitoring tasks remember previous runs and can stop when an end condition is reached.

Building Interactive Sites

ChatGPT Sites lets Work turn an idea or dataset into an interactive website or lightweight app.

You can create a Site by asking Work to build one, or invoke @Sites directly.

Examples include:

  • Dashboards
  • Project trackers
  • Launch calendars
  • Internal portals
  • Interactive reports
  • Lightweight prototypes
  • Shareable data tools

You can preview the Site inside ChatGPT, revise it conversationally, and publish it through a URL.

ChatGPT Work desktop app displaying a completed Sites project: the Hermes Agent Release Radar dashboard with Command Center tab showing release intelligence, commit status, and feature tracking in a dark mode interface.
Custom Sites project: Hermes Agent Release Radar dashboard built with ChatGPT Work.

Sites is currently in public beta.

OpenAI says it’s available on paid plans other than Free and Go, with regional restrictions at launch. Workspace administrators may also control who can create or publish Sites.

For Enterprise workspaces, public publishing is off by default until an administrator enables it.

A useful Site prompt might look like this:

@Sites

Build an interactive launch tracker from the attached spreadsheet.

Requirements:
- A dashboard showing overall completion
- Filters for owner, team, status, and due date
- A timeline view
- A table of blocked items
- A section showing tasks due in the next seven days
- Responsive layout for desktop and mobile
- No public publishing

Use only the attached data. Flag malformed rows and missing owners. Preview the Site for me and stop before publishing it.

The phrase “stop before publishing” belongs in every early Site experiment.

Choosing Sol, Terra, or Luna

ChatGPT Work uses the GPT-5.6 model family.

OpenAI recommends three main models.

GPT-5.6 Sol

Sol is the flagship.

Use it for:

  • Complex, open-ended tasks
  • Difficult research
  • High-value analysis
  • Computer Use
  • Work requiring judgment
  • Polished documents
  • Ambiguous problems with tradeoffs

Sol is the model to reach for when the cost of a weak result is higher than the cost of extra reasoning.

GPT-5.6 Terra

Terra is the everyday workhorse.

Use it for:

  • General professional work
  • Strong reasoning without Sol’s full depth
  • Routine analysis
  • Standard documents
  • Normal tool use
  • Tasks you previously gave a frontier model by default

Terra is the sensible middle lane.

GPT-5.6 Luna

Luna is the fast, lower-cost option.

Use it for:

  • Extraction
  • Classification
  • Transformation
  • Structured summaries
  • Clear repeatable jobs
  • High-volume work where the output contract is already known

Don’t pay Sol prices to alphabetize a clean list.

Reasoning Levels, Max, and Ultra

Higher reasoning can improve difficult work, but it also takes longer and consumes more usage.

OpenAI recommends using the lowest reasoning level that produces the result you need.

  • Light: Quick, tightly scoped work
  • Medium: Everyday tasks needing some planning
  • High or Extra High: Difficult work with several steps, sources, or tradeoffs
  • Max: More reasoning time for one extremely difficult task
  • Ultra: Subagents divide suitable parts of a complex task and work in parallel

Most tasks don’t need Max or Ultra.

Ultra makes sense when the work divides cleanly.

For example, a large competitive analysis could separate into:

  • Product research
  • Pricing analysis
  • Customer segmentation
  • Feature comparison
  • Risk analysis
  • Presentation creation

A small rewrite doesn’t need an agent swarm kicking down six doors at once.

Credits, Usage, and Why Task Scope Matters

ChatGPT Work and Codex share the same agentic usage, credits, and limits.

Regular Chat usage is separate.

The amount consumed depends on:

  • Model
  • Prompt size
  • File size
  • Conversation history
  • Tool results
  • Reasoning level
  • Number of steps
  • Output length
  • Computer Use
  • Image generation
  • How long the agent keeps working

OpenAI says GPT-5.6 usage averages roughly 5 to 40 credits per message, but that’s an average across work with very different shapes.

A tiny extraction job and a two-hour research workflow aren’t the same unit of labor just because each started with one message.

To make usage last longer:

  • Use Luna for clear, repetitive work.
  • Use Terra for normal professional tasks.
  • Reserve Sol for difficult, ambiguous, or high-value work.
  • Provide only relevant files.
  • Limit the source range.
  • State the output length.
  • Separate required work from optional improvements.
  • Avoid dumping a warehouse of context into every task.
  • Start a fresh task when an old thread has become bloated.
  • Don’t use Max or Ultra by reflex.

The best cost control isn’t writing a shorter prompt.

It’s defining a smaller job.

Permissions, Approvals, and Safety

Work is designed to keep you in control of sensitive actions.

Depending on the tool and workspace settings, ChatGPT may ask before:

  • Accessing a new website
  • Using a connected app
  • Sending information
  • Changing a record
  • Submitting a form
  • Making a purchase
  • Deleting data
  • Changing permissions
  • Sharing sensitive information

Workspace administrators can control:

  • Which Plugins are available
  • Which apps can connect
  • Which roles can access them
  • Which actions are allowed
  • When approvals are required
  • Browser and network access
  • Company context
  • Local and cloud policies

Connected apps respect the permissions of the authenticated user. ChatGPT should only be able to reach information that user is already allowed to access.

Still, permission isn’t judgment.

Treat website content as untrusted.

Review the active account before approving an action.

Keep financial, publishing, account, and destructive actions behind explicit approval.

Never paste passwords or secret keys into the conversation.

Put credentials into the actual sign-in interface.

Current ChatGPT Work Limitations

ChatGPT Work arrived with a large surface area, but it isn’t finished.

Here are the limitations worth knowing immediately.

Cloud and desktop Work threads do not sync

Web and mobile Work conversations remain in the cloud.

Desktop Work threads remain on that computer.

Web and mobile cannot directly access local computer files

Local folders and desktop apps require the desktop app.

Scheduled Tasks cannot currently use Project files

A scheduled task created in a Project doesn’t automatically gain access to the Project’s uploaded files.

The Scheduled page is not currently in the desktop app

Manage scheduled tasks on web or mobile.

PowerPoint is not included in the Work desktop file flow at launch

OpenAI offers a separate ChatGPT for PowerPoint experience.

Direct work inside an open Excel workbook uses Codex and the Excel add-in

Work can create spreadsheet files, but the open-workbook integration follows a different path.

The built-in browser has its own browser state

It doesn’t automatically inherit Chrome cookies, open tabs, or signed-in sessions.

Computer Use cannot automate browser file uploads at launch

You may need to handle uploads manually.

Plugins and apps can be unavailable even when Work is available

Access depends on plan, role, region, surface, administrator settings, and the app itself.

Sites is still a public beta

Limits, availability, publishing controls, and features may change.

These are launch seams, not reasons to ignore the product.

They are reasons to design your workflow around the product that exists today instead of the one you assume will exist next month.

Practical ChatGPT Work Ideas

The best Work tasks have a clear result and source material you trust.

1. Build an executive report

Provide spreadsheets, notes, and a previous report template.

Ask Work to reconcile the numbers, explain the largest changes, create charts, and produce the finished report.

2. Create a competitor comparison

Give it a product list and approved source boundaries.

Request a spreadsheet, scorecard, narrative summary, and recommendation.

3. Prepare for a meeting

Connect the relevant email, calendar, CRM, and file sources.

Ask for background, recent activity, unresolved questions, risks, and a meeting brief.

4. Turn research into a presentation

Attach the research pack and a master deck.

Ask Work to preserve the master, build the story, create speaker notes, and flag weak claims.

5. Create a campaign package

Provide the customer research, brand guidelines, launch plan, and product facts.

Request the campaign brief, messaging matrix, channel adaptations, and review checklist.

6. Review a folder of documents

On desktop, grant access to one project folder.

Ask Work to build a summary, decision log, issue list, and clean handoff package without changing the originals.

7. Monitor a changing source

Create a monitoring task for official release notes, a dashboard, a pricing page, or another approved source.

Ask for notifications only when the change is meaningful.

8. Refresh a recurring meeting agenda

Connect the approved team tools.

Have Work gather updates, identify blockers, and update the agenda before the meeting.

9. Build an interactive dashboard

Give Sites a clean dataset and explicit display requirements.

Ask it to preview the result and stop before public publishing.

10. Audit a deliverable

Give Work the output, sources, template, and acceptance criteria.

Ask it to identify missing sections, unsupported claims, broken formulas, inconsistent formatting, and unresolved issues.

Common Mistakes

Asking for activity instead of an outcome

“Research this” is weaker than “create this report using these sources.”

Giving it every file you own

More context can mean more noise, more cost, and more opportunities to use the wrong information.

Hiding the acceptance criteria in your head

Work can’t reliably meet a standard you never stated.

Letting it invent around missing information

Tell it to flag gaps and conflicts.

Forgetting the approval boundary

State what it may do and where it must stop.

Using Sol and Ultra for everything

The biggest hammer isn’t automatically the best screwdriver.

Treating the first output as final

Review the files. Check the formulas. Open the links. Inspect the citations. Read the conclusion.

Building a recurring task before the manual workflow is stable

Run the job manually first.

Fix the prompt.

Then schedule it.

Automation makes a good process faster.

It also gives a bad process a calendar.

The Complete ChatGPT Work Master Prompt

Cyberpunk-style banner reading One Prompt to Rule Them All with a glowing central document icon surrounded by connected tool panels for search, file management, web browsing, scheduling, data dashboards, and workflow automation.
Visual banner for the ChatGPT Work Master Prompt section.
You are completing a substantial task in ChatGPT Work.

GOAL
[Describe the real-world goal.]

FINISHED DELIVERABLES
Create:
1. [Deliverable]
2. [Deliverable]
3. [Deliverable]

AUDIENCE AND USE
The result will be used by:
[Audience]

It will support:
[Decision, meeting, publication, workflow, review, or action]

APPROVED SOURCES
Use:
- [Files]
- [Project context]
- [Connected apps]
- [Plugins or Skills]
- [Approved websites]
- [Local folder]

Do not use:
- [Unapproved sources]
- [Rumors]
- [Old versions]
- [Unverified user-generated content]

WORKFLOW
1. Read all source material first.
2. Extract confirmed facts, requirements, constraints, and unresolved questions.
3. Create a concise plan.
4. Complete the work in logical stages.
5. Preserve the required templates, formulas, branding, filenames, and structures.
6. Cite or link every factual claim that came from an external source.
7. Flag missing or conflicting information instead of guessing.
8. Verify every deliverable before returning it.

OUTPUT REQUIREMENTS
- Format: [Document, spreadsheet, deck, report, Site, folder]
- Length: [Target]
- Required sections: [List]
- Required tables or charts: [List]
- Tone: [Tone]
- File names: [Names]
- Destination: [Where to create or save]

DO NOT
- Invent facts, numbers, dates, quotes, sources, or firsthand experience.
- Modify protected source files.
- Contact people.
- Publish or send content.
- Delete data.
- Change permissions.
- Make purchases.
- Submit information without approval.

STOP AND ASK WHEN
- A required source is missing.
- Approved sources conflict materially.
- An action could be destructive or public.
- A login, payment, permission, or external submission is required.
- The requested result cannot meet the acceptance criteria.

ACCEPTANCE CRITERIA
The task is complete only when:
- Every requested deliverable exists.
- Every required section is present.
- Files open correctly.
- Formulas and calculations have been checked.
- Claims are traceable to approved sources.
- Assumptions are labeled.
- Limitations and unresolved questions are listed.
- No prohibited action was taken.

Return:
1. A concise completion summary
2. Links or paths to every deliverable
3. Verification performed
4. Assumptions made
5. Unresolved issues
6. Actions requiring my approval

Save that template.

Then remove the sections you don’t need for smaller jobs.

Final Takeaway

ChatGPT Work is OpenAI’s attempt to move ChatGPT beyond the answer box.

Chat is still there. Codex is still there.

Work sits between them as the general-purpose execution layer for research, files, tools, deliverables, and repeatable workflows.

The people who get the most out of it won’t be the people who type the longest prompts. They will be the people who define work clearly.

A useful outcome. The right sources. A few hard constraints. A visible quality bar. A line the agent can’t cross without approval.

Start with one job you already know. Run it manually. Watch where Work gets confused. Tighten the sources. Clarify the deliverable. Add the missing acceptance criteria.

Then turn the stable process into a template, Skill, Plugin workflow, or Scheduled Task.

That’s the progression.

Question. Task. System.

ChatGPT Work is the part that finally makes the middle step feel native.

Official Sources

Tony Simons

Reviewed & Written By

Tony Simons

Independent tech reviewer and creator of Tony Reviews Things. 14 years of hands-on testing, software auditing, and workflow automation. I test the gear so you don't waste your money on junk.

Submit a Take

Your email address will not be published. Required fields are marked *