ChatGPT Work: How to Use It, 20 Prompts, and What It Cannot Do

Choose Work for the shape of the task

ChatGPT is still the better starting point for a quick question, brainstorm or short rewrite. Work is intended for a goal that unfolds: research a topic, analyse a set of files, create a report, prepare a spreadsheet, draft a presentation or coordinate a longer deliverable. The distinction is not that Work is always smarter. The distinction is that the task has context, steps, tools and an artifact someone will inspect.

OpenAI's Help Center says Work is available on the web and mobile for eligible paid plans, and in the desktop app when included for the plan and workspace. The exact experience can vary by account, country, administrator and connected app. If Work is missing or greyed out, check the current account and workspace settings instead of assuming the prompt is wrong. Do not promise a feature to a reader based on a different surface. Record where a workflow was run when documenting it.

Write a brief that Work can execute

Start with the outcome, not a vague command such as 'make this better'. State the audience, decision or action the output should support, source material, date boundary, tone, length, destination format and definition of done. Name the files or apps it may use and what is out of scope. Tell it what to do when evidence is missing: mark UNKNOWN, ask a question or leave a placeholder.

A strong brief also states the approval boundary. Work may draft a customer email, but a person sends it. It may prepare a budget, but an owner checks formulas and assumptions. It may create a slide deck, but the presenter approves claims and visuals. It may propose a calendar or task update, but the responsible person confirms dates and recipients. Put these rules in the first message so the workflow does not have to infer authority from an ambiguous goal.

Ask for a plan before execution

For any task with several files or actions, ask Work to return a plan with steps, dependencies, tools it expects to use, risks, required decisions and the evidence it will produce. Check whether it has the right sources and whether any step would change an original file or contact another person. Approve or revise the plan before it runs.

During execution, request checkpoints rather than a single final answer. Each checkpoint should include work completed, sources used, assumptions, unresolved questions, files changed and the next proposed action. This makes it easier to catch a wrong interpretation while the cost of correction is low. If Work pauses for clarification, answer the question with the smallest additional context needed. Do not respond with a blanket permission such as 'do whatever is necessary' when the task includes external side effects.

Use connected apps without losing control

Connections are useful because the best work often depends on the real documents, sheets, messages, calendars or project records. They also increase the consequence of a mistake. Before connecting an app, check the account, data scope and organisation policy. Use a copy or draft area when possible. Ask Work to list the files it actually used and to distinguish a source it read from a source it inferred.

The desktop and cloud experiences can differ. OpenAI documents that Work on web and mobile runs in the cloud, while desktop flows may use local files and desktop apps with permission. Cloud Work chats can sync across surfaces, and messages or task context may be stored in the cloud. Choose the surface and data handling that match the sensitivity of the job. Never paste secrets or unnecessary personal data simply because a connector is available. Access is not the same as permission to redistribute.

Build a research brief with an evidence table

Give Work a focused question, the audience and a cutoff date. Ask for a source plan first, then provide the approved documents or ask it to search where that capability is available. Require a table with claim, source, location, date, confidence, contradiction and next check. Keep direct quotes short and verify them in the original. Ask it to separate what the source says from its interpretation and recommendation.

For current product details, pricing, model names, availability or shutdowns, use the vendor's primary documentation or announcement and record when you checked it. A generated citation is not evidence until you open the link. Do not let Work combine a press release, a user comment and a benchmark into one fact. If the evidence is thin, the correct output may be a list of unanswered questions rather than a confident recommendation.

Create a spreadsheet that can be audited

Describe the decision the spreadsheet should support, then name the sheets, columns, units, periods, formulas and output views. Provide the source rows or attach the workbook. Ask Work to preserve the original in a copy, list all formulas it adds or changes and keep assumptions separate from observed data. Require checks for blank values, duplicate keys, date ranges, currency and totals.

After creation, inspect the workbook itself. Click through formulas, sample a few calculations by hand, check filters and hidden rows, verify charts and read the labels. Ask for a change log but do not rely on the explanation as proof. If the sheet contains payroll, customer, health or financial data, minimise the fields and follow the relevant policy. A polished table can still have a wrong unit, a shifted range or a silent assumption. The person accountable for the decision owns the final review.

Draft a presentation with a real argument

Tell Work the single takeaway, audience, speaking time, decision required and evidence available. Attach a reference deck if the structure or brand matters, and say what must remain unchanged. Ask for a narrative outline before slide production. Each slide should have a point title, one job, source notes and a clear visual need. Mark where a real chart, number, quote or screenshot is required rather than allowing filler.

OpenAI's documented file workflow supports creating or editing documents, spreadsheets and presentations, but availability depends on plan, workspace, file type and surface. It also notes that generated slides may not match a template perfectly and that the output needs review. Duplicate the deck first. Inspect editable objects, speaker notes, citations, page order, overflow, contrast and any generated images. The presenter approves the story and every claim before the deck is shared.

Use Work for repeatable operations

A recurring workflow should have a stable input checklist, a fixed output schema and an exception path. Examples include a weekly project brief, a monthly variance memo, a meeting preparation pack or a support backlog review. Ask Work to identify missing inputs, preserve links to source records, flag changes since the last period and keep recommendations separate from status facts.

Do not turn a repeatable task into an unattended publication pipeline until you have reviewed several real runs. Compare the result with the source, record false positives and refine the brief. Make the owner, review time and escalation route visible. If a recurring task can send messages or alter systems, require explicit approval for the action, not only approval of the template. Automation is successful when it makes responsibility clearer, not when it makes an action impossible to inspect.

Use the 20 prompt patterns

These compact prompts are starting points, not guarantees:

1. 'Turn these notes into a decision brief with options, trade-offs and open questions.' 2. 'Build a source matrix; cite every claim by file and page.' 3. 'Compare these two proposals using the same five criteria.' 4. 'Extract actions, owners and due dates; mark ambiguous owners.' 5. 'Draft a meeting agenda from this decision log.' 6. 'Create a weekly status report; separate facts, risks and asks.' 7. 'Prepare an executive summary followed by evidence.' 8. 'Turn this policy into a staff checklist without changing obligations.' 9. 'Find contradictions between these documents.' 10. 'Create a spreadsheet schema and validation rules before filling it.' 11. 'Audit formulas and report suspicious ranges.' 12. 'Turn this data into a chart brief; do not invent a trend.' 13. 'Create a 10-slide outline with a single takeaway.' 14. 'Adapt this deck to a new audience; preserve cited claims.' 15. 'Draft three email options; do not send.' 16. 'Prepare a research plan with primary sources first.' 17. 'Review this process for privacy and permission risks.' 18. 'Convert this project goal into checkpoints and a definition of done.' 19. 'List what you could not verify.' 20. 'Before finalising, run a source, privacy, formatting and completeness audit.'

For each, add the actual source pack, audience and approval boundary.

Manage memory, Projects and context carefully

A long-running project benefits from a stable brief, a glossary, a decision log and a source register. Keep those artifacts current rather than asking Work to remember an instruction that was never written down. Project settings, workspace controls and the surface you are using affect what context is available. The supplied queue note's claim about project-only memory should not be treated as universal: check the current OpenAI settings and Help Center for your account.

When a task crosses projects or teams, restate the relevant context and confirm the audience. Do not assume a previous chat's permission applies to a new file or recipient. Remove stale instructions and mark superseded decisions. If Work cites an earlier message, check that the message is still authoritative. A context pack is safer than hidden memory because another person can inspect, update and approve it.

Review the artifact, not just the conversation

The final answer can sound finished while the file remains wrong. Open the document, spreadsheet or presentation and inspect the actual output. Check titles, dates, links, formulas, page breaks, charts, citations, permissions and redactions. Compare the output against the brief and make a claim ledger for anything consequential. Ask Work to list omissions and assumptions, then resolve them yourself or assign an owner.

Use a two-pass review. First check substance: is the decision supported, are calculations reproducible, are instructions complete and are risks visible? Then check communication: can the intended audience understand it, are headings useful, are tables readable and is the call to action clear? Keep the original and the generated version. Record the reviewer and date. This is especially important when Work has edited an existing artifact, because a chat transcript does not show every visual or formula-level change.

What ChatGPT Work cannot do for you

Work cannot make an unverified source true, establish that a legal or medical conclusion is safe, guarantee a spreadsheet's formulas, make a presentation persuasive to every audience or take responsibility for a message sent under your name. It may lack the current file, misunderstand a connector, omit a qualification or follow a stale instruction. A plan is not completion, and completion is not correctness.

Keep high-impact judgment with the accountable person. For financial, employment, security, safety, health or legal work, add qualified review and current primary sources. For external communication, require a named sender to read and approve. For destructive changes, use a copy, backup and explicit confirmation. For privacy, minimise inputs and inspect where the output will be stored. The best Work prompt makes these boundaries visible before the first tool call.

A reusable ChatGPT Work brief

Copy this and customise it:

'Outcome: [specific deliverable] for [audience] supporting [decision/action]. Source material: [files, links, connected apps]. Date boundary: [date]. Constraints: [policy, tone, length, format, budget]. First return a plan with steps, dependencies, risks, required permissions and checkpoints; wait for my approval. Work only from the supplied or explicitly accessed sources. Label facts, inferences, recommendations and UNKNOWN items separately. Do not invent citations, numbers, quotes, test results or product capabilities. Do not send, publish, delete, purchase, change permissions or edit an original without explicit approval. At each checkpoint report sources used, assumptions, files changed, checks run and unresolved questions. Final output must include [schema] plus a source register, change log and review checklist.

Before I share or rely on the result, I will inspect the artifact and approve consequential claims and actions.'

This brief works because it defines the job, the evidence boundary, the stop conditions and the handoff. It does not remove the need for judgment.