AI Prompts for Work: How to Write Them, With Examples

Write the prompt like a brief for a smart new colleague who knows nothing about your situation. Say what you need (the task), give the facts it can't know (the context), say who will read it (the audience), describe the shape of the result (the format), and add any rules, such as "use only these facts". Then read the answer, and ask for changes in plain words.

• The AI only knows what you type. Leave out the facts and it fills the gaps with guesses or [PLACEHOLDERS].

• Five parts cover most office tasks: task, context, audience, format and rules.

• "Use only the information above" is the most useful rule. It stops the model from inventing details.

• Your first prompt is a draft. A follow-up such as "shorter, and put the decision first" is normal.

• You still check the result. Even a good prompt can produce a wrong date or an extra promise.

A chat assistant writes the most likely answer to the words in front of it. If you write "Write an email about a delayed order", the most likely answer is a template that fits every company, because the model has no order, no customer and no new date to work with. Every fact you add narrows the answer toward the one you need.

The five parts of a work prompt:

The vendors teach the same idea with different labels. Microsoft's Copilot prompt guide uses goal, context, source and expectations, and says only a clear goal is required. Google's tips for Gemini say to use natural language, give context and break complex tasks into separate prompts. OpenAI's ChatGPT help article says to be clear and specific and to refine the prompt after reading the answer. Anthropic's Claude prompting guide suggests a simple test: if a colleague with no background would be confused by your prompt, the model will be too.

You don't need all five parts every time. A quick question needs only the task. The more the result matters, and the more it depends on your facts, the more parts you add.

AI should not make decisions about people on its own. If you use it anywhere near hiring, performance reviews, credit, health or legal matters, involve HR or legal first and keep a person responsible for the decision. Before you paste anything confidential, read is it safe to use ChatGPT at work?.

Each example shows a weak prompt, a better one, and the model's real answers, trimmed with "…". The answers come from OpenAI's gpt-5-mini, a small model from the family behind ChatGPT. Names, orders and numbers are made up. ChatGPT, Claude, Gemini and Copilot will word their answers differently, but the difference between the weak and the better prompt holds across tools.

1. Write an email to a customer

The weak prompt:

Write an email to a customer about their delayed order.

Subject: Update on Your Order [#ORDER_NUMBER] — Delivery Delay & Next Steps

Hello [Customer Name],

I’m writing with an update about your order #[ORDER_NUMBER] placed on [ORDER_DATE]. I’m very sorry to let you know that your shipment of [ITEM(S)] has been delayed.
…
How we’re making it right
We understand this is inconvenient. As an apology we’d like to offer:
- [Option A — e.g., X% off your next order] OR
- [Option B — e.g., refund of shipping charges of $X] OR
- [Option C — e.g., expedited shipping at no cost]
…

That is a template with more than a dozen blanks, and it offers a discount you may not be allowed to give. The better prompt uses all five parts:

Task: Write an email to a customer whose order will arrive late.

Context: The customer is Priya Shah, office manager at a dental practice. She ordered 40 ergonomic chairs (order #44817) for delivery on May 6. Our supplier shipped late, and the new delivery date is May 14. We will refund the $180 shipping fee. She can cancel without a fee if May 14 doesn't work for her.

Audience and tone: She is a long-time customer. Be warm and direct. Apologize once. Don't blame the supplier.

Format: A subject line, then the email. Under 120 words. End with one clear next step for her.

Rules: Don't promise anything that isn't listed above.

Subject: Update on Order #44817 — New Delivery Date May 14

Hi Priya,

I’m sorry to let you know your order of 40 ergonomic chairs (Order #44817), originally scheduled for May 6, will arrive on May 14. We will refund the $180 shipping fee to your original payment method. If the May 14 delivery doesn’t work for your practice, you may cancel the order without any fee.

Please reply to confirm whether you’d like us to proceed with the May 14 delivery or cancel the order.

Thank you for your understanding, and I appreciate your continued trust.

Nearly ready to send. One phrase wasn't in the facts: "to your original payment method". It is probably true, but check it before you send. That is why the rule "don't promise anything not listed" matters, and why you still read the result.

2. Summarize a long document

The memo below is about 190 words. Both prompts paste it in full.

Subject: Updated travel and expense policy (effective July 1)
From: Finance Operations

Starting July 1, all travel must be booked through the TripDesk portal. Bookings made elsewhere will not be reimbursed unless Finance approves an exception in writing before the trip.

Hotel limits now depend on the city. In New York, San Francisco and London the nightly cap is $325. In Chicago, Boston, Seattle and Toronto it is $225. Everywhere else it is $175. Rates above the cap need manager approval in TripDesk before you book.

Meals: the daily meal allowance rises from $60 to $75. Alcohol is no longer reimbursable, including at client dinners. Client entertainment over $500 needs VP approval.

Receipts: submit receipts within 30 days of the expense. Expenses submitted more than 60 days after the expense will not be reimbursed. Receipts under $25 are no longer required, but every expense still needs a short description.

Mileage stays at 70 cents per mile.

Corporate cards: employees who travel more than four times a year will get a corporate card. Personal cards can still be used, but reimbursement can take up to two pay cycles.

Questions: travel@brightline.example

With "Summarize this." the model returned seven bullets that restate every paragraph in a slightly shorter form:

Here’s a concise summary of the updated travel & expense policy (effective July 1):

- All travel must be booked via the TripDesk portal; outside bookings require prior written Finance approval to be reimbursed.
- Hotel nightly caps by city: New York/San Francisco/London $325; Chicago/Boston/Seattle/Toronto $225; all other locations $175. …
…
- Mileage reimbursement remains 70¢ per mile.
…
- Questions: travel@brightline.example

Accurate, but not much shorter, and not written for anyone. The better prompt says who reads it and what they care about:

Task: Summarize the policy memo below for our sales reps.

Context: The reps travel two or three times a month. They care about what changes for them and what they must do differently.

Format: At most 5 bullets, each under 20 words, most important first. Then one line: "Effective date: ...".

Rules: Use only the memo. Don't add advice. If a rule is unclear, add a final line that starts with "Unclear:".

Memo:
"""
[paste the memo here]
"""

- All travel must be booked via TripDesk; bookings elsewhere need prior written Finance exception.
- You'll receive a corporate card if you travel more than four times yearly; personal cards may delay reimbursement.
- Hotel nightly caps: NYC/SF/London $325; Chicago/Boston/Seattle/Toronto $225; elsewhere $175; over-cap needs manager TripDesk approval.
- Meals allowance $75/day; alcohol not reimbursed; client entertainment over $500 requires VP approval.
- Submit receipts within 30 days; expenses >60 days won't be reimbursed; receipts under $25 not required.

Effective date: July 1.

The model now chose what matters to someone who travels every month and dropped the rest. The triple quotes around the memo show the model where your text starts and ends. For more ways to keep the key facts in a summary, see Chain of Density prompting.

3. Turn meeting notes into action items

mtg 5/12 - Q3 webinar planning
attendees: Dana, Luis, Mei, Sam (joined late)
- topic: AI for finance teams? Luis likes it, Mei worried too broad. agreed: "Closing the books faster with AI", target CFOs at mid-size companies
- date: Aug 20 or Aug 27. Dana to check speaker availability (Raj from Ledgerly)
- Mei: landing page draft, needs final topic first
- promo budget $4k. Sam ok with LinkedIn ads but wants last webinar numbers before committing more
- Luis will pull last webinar stats
- open: do we record it? need to ask Raj if ok
- next mtg next Tues

With "Clean up these meeting notes." the recap looked tidy, but it added things nobody said:

…
Action items
- Dana: Confirm Raj's availability for Aug 20 or Aug 27; ask Raj whether recording is allowed. (Owner: Dana)
…
- Mei: Hold landing page draft pending any changes; will finalize after speaker confirmation. (Owner: Mei)
- Sam: Review past webinar numbers once provided and confirm promo channel/budget allocation. (Owner: Sam)

Next meeting
- Next Tuesday (5/16).

The notes don't say Dana will ask about recording, and they give Sam no task. And "5/16" is a guess: the notes have no year, and May 16 is a Tuesday only in some years. Send that to four people and someone books the wrong day. The better prompt:

Task: Turn these meeting notes into a short recap I can send to the attendees.

Format:
1. Decisions (bullets)
2. Action items as a table: Task | Owner | Due
3. Open questions (bullets)
Keep it under 150 words.

Rules: Use only what is in the notes. If an owner or due date is not stated, write "not stated". Do not add tasks.

Notes:
"""
[paste the notes here]
"""

…
Action items
Task | Owner | Due
Check Raj (Ledgerly) speaker availability | Dana | not stated
Prepare landing page draft (awaiting final topic) | Mei | not stated
Pull last webinar stats | Luis | not stated
Ask Raj if recording is OK | not stated | not stated

Open questions
- Which date (Aug 20 or Aug 27) will be confirmed?
- Do we record the webinar (need Raj's permission)?

"Not stated" is useful: it shows you which gaps to fill before you send. Read the decisions too. This version listed "Proposed dates: Aug 20 or Aug 27" as a decision, which it isn't yet.

4. Write a status report

"Write a status update for my website redesign project." produced almost 500 words about work that never happened:

Summary
- Project on track overall. Core design and front-end build are largely complete; back-end integration and QA are in active progress.
…
- Production launch: on track for mid-November if current pace and dependency resolutions hold.
…
- Lighthouse performance baseline (pre-optimization): Home page 68; …

With no facts, the model invents plausible ones, including a launch month and a performance score. Give it your facts and your reader:

Task: Write this week's status update on our website redesign for our VP of Marketing.

Context (week of June 9):
- Done: homepage and pricing page designs approved by marketing on June 5.
- In progress: product pages, 6 of 14 built.
- Problem: the product photo shoot moved from June 12 to June 26 because the studio double-booked. Product pages can't go live without the new photos.
- The launch date was July 1. July 15 is realistic now.
- We need a decision by June 13: launch July 1 with the old photos, or wait until July 15.

Audience: She reads updates on her phone between meetings and wants to see any decision she must make first.

Format: Subject line. Overall status in one word (On track, At risk or Off track). Then "Decision needed", then Done, Next and Risks with at most 3 bullets each. Under 150 words.

Rules: Use only these facts.

Subject: Website redesign status — week of June 9

Overall status: At risk

Decision needed (by June 13): Launch July 1 using old product photos, or delay launch to July 15 to wait for new photos.

Done:
- Homepage and pricing page designs approved by Marketing on June 5.

Next:
- Continue building product pages (6 of 14 completed).
…
Risks:
- Product photo shoot moved from June 12 to June 26 (studio double-booked).
…

The decision is at the top, where a busy reader sees it, and every line traces back to a fact you gave.

5. Analyze a table of numbers

Month,Billing,Login issues,Shipping,Product questions,Total
Jan,310,120,450,280,1160
Feb,295,115,430,300,1140
Mar,320,130,470,290,1210
Apr,305,410,480,295,1490
May,315,520,460,310,1605
Jun,300,480,455,305,1540

"Analyze this data." returned a long report with shares of the six-month total, a list of generic recommendations and an offer to make charts. It did spot the jump in login issues, but you have to dig for it. Ask the question you actually have:

Task: Explain what drove the increase in support tickets from Q1 (Jan-Mar) to Q2 (Apr-Jun).

Context: We sell accounting software to small businesses. We launched a new login page on April 2.

Format: First the answer in one sentence. Then a table: Category | Q1 total | Q2 total | Change. Then at most 3 bullets on what to check next.

Rules: Show the numbers you used. Label any cause you suggest as a hypothesis. Use only this data.

Ticket counts by month:
[paste the table here]

Answer (one sentence):
The increase in support tickets from Q1 to Q2 (+1,125 tickets) was driven overwhelmingly by a large rise in Login issues after April (Login issues +1,045), which coincides with the new login page launch (hypothesis).

Table (numbers used shown):
Category | Q1 total | Q2 total | Change
Billing | 925 (310+295+320) | 920 (305+315+300) | -5
Login issues | 365 (120+115+130) | 1,410 (410+520+480) | +1,045
Shipping | 1,350 (450+430+470) | 1,395 (480+460+455) | +45
Product questions | 870 (280+300+290) | 910 (295+310+305) | +40
Total | 3,510 (1160+1140+1210) | 4,635 (1490+1605+1540) | +1,125
…

"Show the numbers you used" lets you check every sum in a minute with a spreadsheet. These are correct, but don't assume that. "Label any cause as a hypothesis" keeps the model from presenting a guess as a finding.

6. Let the AI ask you questions first

When you don't know what to include, ask the model to interview you before it writes:

I need a job posting for a part-time customer support agent at our company. Before you write anything, ask me the questions you need answered to write a good posting. Ask at most 5 questions, all at once, and wait for my answers.

1. Company name, brief description (mission, product/service), and any tone/branding preferences for the posting (e.g., formal, friendly, startupy)?
2. Is this role remote, onsite, or hybrid? If onsite/hybrid, what city/location and expected schedule or timezone coverage?
3. Key responsibilities and typical weekly hours or shift times …?
4. Required and preferred qualifications/skills …
5. Compensation range, benefits/perks for part-time employees, and application instructions …

Answer the questions in one message, and the model has the context it needs. Have HR review any job posting before it goes out.

Paste this into any AI tool, fill in the brackets, and delete the parts you don't need.

Task: [What you need, starting with a verb: write, summarize, turn into, compare, explain]

Context: [The facts the AI can't know. Paste notes, numbers or a document between triple quotes.]
"""
[paste here]
"""

Audience and tone: [Who will read it, what they care about, how it should sound]

Format: [Length, sections, bullets or a table with named columns, subject line]

Rules: Use only the information above. If something is missing, write "not stated" or ask me. [Anything else to avoid]

Then improve the answer with follow-ups instead of starting over:

Shorter. Keep only what my manager needs to decide.
What did you assume that isn't in my notes?
Rewrite this for someone who has never heard of the project.
Give me three subject lines to choose from.
Put this in a table with the columns Task, Owner and Due.

If your team writes the same kind of prompt every week, save the filled-in template in a shared document. Copilot, ChatGPT, Claude and Gemini also let you save custom instructions or reusable prompts, though the feature names differ by tool and plan.

What is the best prompt formula for work?

Task, context, audience, format and rules. Say what you need, give the facts, name the reader, describe the shape of the result, and add "use only the information above". For quick questions, the task alone is enough. Add the other parts when the result depends on your facts or will be sent to someone.

Do the same prompts work in ChatGPT, Claude, Gemini and Copilot?

Yes. The five parts work in every major assistant, because they all answer from the words you give them. The wording, length and style of the answers will differ between tools and between model versions. Some assistants can also read your work files and email if your company has connected them, which changes what context you need to paste.

How long should a work prompt be?

As long as the facts require. A rewrite request can be one sentence. An email that depends on an order number, dates and a refund needs those facts, which is often 80 to 150 words. Long is fine when every line is a fact or an instruction the model needs.

How do I stop ChatGPT from making things up in my drafts?

Give it the facts, and add a rule: "Use only the information above. If something is missing, write 'not stated'." Then read the result and compare names, numbers, dates and promises with your facts. A good prompt reduces invented details. It doesn't remove them. See how to check AI answers.

Should I tell the AI to "act as" an expert?

It can help with tone and vocabulary, for example "write as an experienced HR coordinator". It doesn't add facts the model lacks, so it is no substitute for context. The guide to role prompting explains when a persona helps.

Can AI do the task for me, not just write it?

A chat assistant drafts text. It doesn't send the email, update the CRM or book the meeting unless your company has connected it to those systems. Software that takes actions in other systems is called an AI agent, and it needs approvals and limits that a chat window doesn't. See what is an AI agent?.

• Keep company data safe while you prompt: is it safe to use ChatGPT at work?

• Catch wrong facts before you send: how to check AI answers

• Learn the parts of a prompt in more depth: what is prompt engineering?

• Write prompts the model can't misread: clear asks, roles, constraints and format

• Show the model a sample of what you want: few-shot prompting

• Control tone and style across a team: write a style guide prompt