The beginner's guide to ChatGPT at work
Someone sends you a spreadsheet export and asks what changed this month. ChatGPT can tell you in a minute; the useful part is checking that minute against the file.
What does your job involve?
Nine ChatGPT jobs that go beyond a text draft: images, files, research, projects and connected apps. Each card says where to start, what your plan needs, a prompt to use and the check that matters.
Open a task to see where to start, the access to check, a prompt to use and what to check in the result.
Create a presentation from a report ChatGPT Work → presentation file
Move from a long report to a deck people can edit and question.
- Where to start
- Turn on Work in the switcher, select + › Add photos & files to attach the report, and ask for a PowerPoint file. Open the file in PowerPoint.
- Access to check
- Work and file creation in your workspace; a text outline alone is not a deck. Use a report you're allowed to share.
- Bring this
- A checked report, audience, decision and intended slide count.
Try this in ChatGPT
Create an editable six-slide PowerPoint from this report for [audience]. Lead with [decision]. Use only facts from the report, keep proposals labelled and put source references in the speaker notes. Include a limitations slide. Return the file for review; do not publish or email it.Why this request helps: The requested file format and decision give the task a destination; source rules prevent a convincing deck from inventing its evidence.
Check before using it: Open every slide; check figures, source references, legibility and editable text.
If this route is missing: Ask Chat for a sourced slide outline, then build it in your usual presentation editor.
Generate a campaign image ChatGPT → image generation
Create a candidate visual rather than another paragraph of campaign ideas.
- Where to start
- In a new chat, describe the image you want. To use a reference, attach it with + › Add photos & files.
- Access to check
- Image generation in your plan, your usage allowance, and permission to use any reference image.
- Bring this
- A campaign purpose, visual style, destination dimensions and exclusions.
Try this in ChatGPT
Generate one image for [campaign purpose], in [style], with [main subject]. Leave a quiet region for editable headline text that I will add later. Use [palette]. No logos, invented product claims or lettering. Make a [destination shape] candidate for review.Why this request helps: Purpose, composition and exclusions are more useful than an adjective such as stunning. Keeping factual text editable makes corrections practical.
Check before using it: Look for impossible joins, extra limbs, stray lettering and unsupported product details. Test the real layout.
If this route is missing: Ask for a written visual brief and take it to your usual image editor.
Improve an informal headshot ChatGPT → uploaded-image edit
Improve presentation while keeping the person recognisable.
- Where to start
- In a new chat, attach the photo with + › Add photos & files and ask for one restrained edit.
- Access to check
- Image editing in your plan, and the person's agreement to use their photo this way.
- Bring this
- A readable original and the intended profile or page crop.
Try this in ChatGPT
Edit this photo for a professional profile. Keep the same face, apparent age, expression, hair, clothes and accessories. Improve the light so it is soft and even, and simplify the background. Keep the full hairline and the same crop of the body. No new hands or outfit. Make one candidate for comparison.Why this request helps: The original supplies appearance; the improvement brief names the limited changes rather than inviting a makeover.
Check before using it: Compare source and result, including glasses, hair, expression and crop. Reject changed appearance.
If this route is missing: Ask for a practical retake plan, then use a better original photo.
Analyse a spreadsheet export ChatGPT Chat → file analysis
Find what changed and understand how the figures were calculated.
- Where to start
- In a new chat, attach the CSV or workbook with + › Add photos & files and name the sheet and period you mean.
- Access to check
- File uploads in your workspace; analysing a copy is different from editing the live spreadsheet.
- Bring this
- The export, column definitions, currency, date range and question.
Try this in ChatGPT
Analyse [sheet or file] for [question]. Define the measure, show the row count and any excluded rows, and calculate totals from the data in the file. Separate observations from possible causes. Return a small table and show the sum for one total so I can check it. Do not fill missing figures with guesses.Why this request helps: Naming units and exclusions makes an attractive summary auditable. A cause is a separate claim from a calculated change.
Check before using it: Recalculate one total and look at duplicate, missing and zero-value rows in the file.
If this route is missing: Paste a small table and ask for a limited explanation; calculate the totals yourself.
Build an editable scenario workbook ChatGPT Work → spreadsheet file
Get a reusable model that can be checked as assumptions change.
- Where to start
- Turn on Work in the switcher and ask for an .xlsx file with named input cells and working formulas.
- Access to check
- Work and spreadsheet creation in your workspace; check the downloaded file, not the preview.
- Bring this
- A simple rule, checked inputs, units and two known-answer cases.
Try this in ChatGPT
Create an editable Excel workbook for [scenario]. Separate inputs, formulas and outputs; include the stated units. Use formulas linked to input cells, flag missing required values and include these known-answer cases: [cases]. Return the file and explain the checks you ran. Do not hard-code the sample total.Why this request helps: A working model needs changing inputs to reach the result through formulas. A screenshot or table cannot establish that.
Check before using it: Change an input in Excel, check formula ranges and test the normal, zero and missing cases.
If this route is missing: Ask for the formula structure, then build it yourself in a copy of the spreadsheet.
Research a current work question ChatGPT → web search
Answer a question whose facts may have changed.
- Where to start
- In a new chat, ask for current official sources. For a larger investigation, use deep research.
- Access to check
- Search in your workspace; deep research allowances vary by plan.
- Bring this
- A specific question, geography, date and evidence standard.
Try this in ChatGPT
Research [question] for [market/date]. Start with official sources. For each finding, give the supporting link and distinguish current fact, announced change and unresolved question. Compare evidence where sources disagree. Do not infer our internal policy from public pages.Why this request helps: Scope and source standards turn a broad question into inspectable research.
Check before using it: Open the important links; check the dates and whether the page supports the claim.
If this route is missing: Open official sources yourself, paste the excerpts and ask for a comparison.
Find information in connected work apps ChatGPT → Plugins tab
Use existing work context when the approved connection supports it.
- Where to start
- Open the Plugins tab, find the service, check what it can read and change, and install it. In a new chat, type @ and the plugin name, then ask one narrow question.
- Access to check
- A plugin your workspace allows, signed in to the right service account.
- Bring this
- A specific project, date range and named source system.
Try this in ChatGPT
In [connected service], find [specific records] for [project and date range]. Return a brief with direct source references and note any records you could not open or find. Search only; do not send messages, update records or change permissions.Why this request helps: A connected-service request needs both a source scope and an action boundary.
Check before using it: Open the cited records in the original service and check which account the plugin used.
If this route is missing: Attach an export instead, and treat it as a snapshot rather than a live connection.
Read a screenshot or chart ChatGPT → image input
Get help understanding a visual without pretending it contains all the evidence.
- Where to start
- In a new chat, attach the screenshot with + › Add photos & files and ask about what is visible.
- Access to check
- Image input in your plan; a picture may leave out rows or screens the question depends on.
- Bring this
- A readable screenshot, the question and relevant definitions.
Try this in ChatGPT
Explain what is visible in this screenshot for [question]. Separate readable labels from uncertainty. If a figure is unclear, say so. Do not infer hidden rows, missing screens or the cause of a trend. Tell me which original data would be needed to verify the conclusion.Why this request helps: An image supplies visible context, not the complete system behind it.
Check before using it: Compare labels and values with the original; supply data for calculations.
If this route is missing: Paste a text extract and the underlying table instead.
Keep a recurring project together ChatGPT → project
Reuse a useful method without recycling an outdated conclusion.
- Where to start
- Open Projects, create or choose the project, add current files to its Sources and write short project instructions. Start each new chat from inside the project.
- Access to check
- Projects in your plan; check who else is a member and whether the files are current.
- Bring this
- A recurring job, stable format and current source set.
Try this in ChatGPT
For this project, use [headings] and keep tentative dates tentative. For each new brief, use the current files I give you and list missing facts. The previous output is a formatting example, not proof that its facts are still current.Why this request helps: Stable instructions reduce repetition; changing facts still need fresh evidence.
Check before using it: Check which files and instructions are in context before the next task.
If this route is missing: Keep the same instruction in a document and paste fresh facts for each task.
Know what you have
ChatGPT: the feature map
Use this map to distinguish a reply, an editable output and an action in another app. Availability depends on the signed-in account, enabled features and the specific operation.
| Feature | Useful for | Where | Access / limit |
|---|---|---|---|
| Chat | Drafts, explanations and revisions | New chat | Your work account; limits vary by plan |
| Work | Multi-step research or finished deliverables | Work switch beside Chat | Plan and workspace; open the finished file |
| Search / research | Current public evidence | Ask in Chat, or use deep research | Workspace settings and plan allowances |
| Image generation / editing | Marketing visuals and portrait candidates | Describe the image, or attach a reference | Image allowance; rights to any reference |
| Files | Analysis, documents, slides and spreadsheets | + › Add photos & files; ask for the format | Copies of files, not live Office editing |
| Projects | Recurring work and shared context | Projects | Membership, instructions and current files |
| Plugins | Approved connected data or actions | Plugins tab; @ in a chat | Service sign-in and what each action can change |
| Codex | Software projects and technical change review | Codex | Separate technical workflow and project access |
Find out what your account can use
- Sign in to the account your organisation uses for work and check the workspace name in the account menu before adding work material.
- Use Chat for a conversation and turn on Work for a task with several steps or a finished file.
- Look for + › Add photos & files, image generation and the Plugins tab. If one is missing, check your workspace settings or ask your administrator.
- Before installing a plugin, check which service account it uses and what it can change. A connected app does not mean every file in the organisation is reachable.
If a feature is unavailable, use the text-only alternative in the task card. Do not move confidential work to a personal account to get a different button.
Ask your administrator:
Which ChatGPT account and workspace should I use for work? Please confirm whether our workspace has Work, file uploads and the Plugins tab, which information I may share, and who agrees new access. I want to try [specific task and destination] with the access we already have.What you'll make
a short “what changed” summary of a small spreadsheet export, with one total you have checked yourself.
What you'll need
ChatGPT in the account your organisation uses for work, and a small CSV or Excel export you’re allowed to share with it. To practise first, copy the fictional file in the worked example. About ten minutes.
Example result.
Total courier spend rose from £12,745 in September to £13,530 in October (+£785, 6.2%). York rose most: +£635 (19.2%). Hull fell by £180 (6.2%). Leeds, Selby and Ripon each moved by less than £300. October total: 4,385 + 2,710 + 3,940 + 1,475 + 1,020 = 13,530.
1. Open a new chat in your work account
Sign in and check the account menu shows your organisation’s workspace, not a personal account. Start a New chat. Use Chat for a question like this one; ChatGPT’s Work switch is for larger tasks that end in a finished file, such as a deck or workbook.
Why this works: the workspace decides what happens to the file you upload, so it’s worth two seconds before the first upload rather than after.
What you should see: an empty message box with a + button on the left.
2. Attach the export and ask what changed
Select + in the message box, then Add photos & files, and choose the export. (For a very small table you can paste it instead.) Then ask:
This file shows [measure] by [row label] for two periods.
Tell me what changed between them in five lines or fewer:
the overall total for each period and the change, then the
biggest rise and the biggest fall in money and percent.
Show the sum you used for the newer total so I can check it.
Describe only what the file shows. Do not suggest reasons.[measure] is what the numbers are, such as “courier spend in pounds”. [row label] is what each row is, such as “depot”.
Why this works: naming the measure stops it guessing units, and asking for the sum turns your check into reading one line rather than redoing the analysis.
What you should see: a short answer, often with a small table, and a written-out sum for the newer total.
3. Check one total yourself
Open the file in Excel or Sheets and add up the newer column, or use a calculator on the sum ChatGPT showed. Then check the biggest rise against its two rows.
Why this works: if the total and the biggest movement are right, the rest of a five-line summary usually is. If either is wrong, nothing else should be trusted until it’s fixed.
What you should see: your total matches the answer to the penny.
Check the result
- The two period totals match your own sums.
- The biggest rise and fall are the rows you would pick from the file.
- Percentages use the older period as the base.
- Every sentence describes the file; none explains why.
The usual failure: a reason appears that isn’t in the file. “York rose because of the new same-day service” sounds helpful, and nobody told ChatGPT about a same-day service. Spot it by asking of each sentence: which cells say this?
Remove every reason or cause that is not in the file.
Keep only the totals, changes and percentages, and show
the sum for each total so I can check it.Worked example: courier costs by depot
Fictional example. A regional manager attached this export:
depot,september_gbp,october_gbp
Leeds,4120,4385
Hull,2890,2710
York,3305,3940
Selby,1450,1475
Ripon,980,1020The first answer gave the right totals, then added: “York’s 19% jump likely reflects the new same-day service launched in October, while Hull’s fall suggests lower demand.”
- The numbers were right. 4,120 + 2,890 + 3,305 + 1,450 + 980 = 12,745; the October column sums to 13,530, a rise of £785 (6.2%). York’s +£635 is 635 ÷ 3,305 = 19.2%.
- The reasons were invented. The file has no column for services or demand. The repair prompt removed both sentences, leaving the summary at the top of this page.
- The check took under a minute. Adding one column and one row pair was enough to trust the five lines, and to know the “why” is a question for the York depot, not for ChatGPT.
Task kit for this job: Download (ZIP) · Reusable instructions
Go deeper
If your team uses a custom GPT, plan for plugins
OpenAI has scheduled custom GPTs to retire on 11 December 2026. Their workflows move to plugins, which can bundle skills (reusable instructions) and apps (connections to other services). If you only use someone else’s GPT, look for its creator’s notice: access to the old GPT doesn’t carry over to the plugin automatically. If you built one, migrate it, then test it on a real task before the date. Custom actions don’t transfer and need rebuilding. Plugins live in the Plugins tab; type @ in a chat to call one by name. Custom GPT retirement FAQ.
Use files, projects and connected apps when they earn their place
A file gives ChatGPT evidence for one chat. A project keeps related chats, files and instructions together, so you stop re-attaching the same export every month. A plugin adds access to another service, and possibly the ability to change things there. Check each separately: what it can read, what it can change, and whose account it uses. A polished answer from last month is not this month’s source. Walkthroughs: Projects, connected files and editable outputs.
Ask for a finished file with Work
When the result is a file rather than an answer, turn on Work in the switcher and ask for the format by name: “an editable six-slide PowerPoint” or “an Excel workbook with formulas linked to input cells”. Then open the file itself. A completion message is not a content check: click through every slide, or change one input and see whether the totals move. Next steps: a presentation or a workbook. For plan choice, see Choose ChatGPT or Claude for your business.