Writing effective prompts fits in one sentence: tell the AI what role it plays, what the context is, what task to do, what format to return, whether there is an example to follow, and what it must avoid. Most of the "generic, useless" answers office workers complain about are not caused by a weak tool but by a one-line request such as "write an apology email to a customer". This lesson teaches the 6-part formula, compares bad and good prompts for everyday office work in Vietnam, and shows how to fact-check results and keep company data safe.
By the end of this lesson, you will be able to:
- Write a prompt with all 6 parts (role – context – task – format – example – constraints) in about 3 minutes.
- Fix 5 common bad prompts (email, meeting minutes summary, comparison table, Excel formula, customer service script) so the result is usable straight away.
- Break a big task into 3–5 small steps and make the AI ask you questions when information is missing.
- Verify figures, citations and calculations before sending anything out.
- Build a personal prompt library and know which data must never be pasted into a public AI tool.
Want to learn to apply AI at work properly? Prompt writing is step 2 of the 8-step roadmap — after understanding how generative AI works and before using AI for documents, spreadsheets, meetings and automation. See the 8-step AI-at-work specialist roadmap.
What is a prompt, and why do people get such different results?
A prompt is everything you send to a chat-based AI tool such as ChatGPT, Gemini, Claude or Copilot: the question, the request, attached documents and examples. A language model does not "get what you mean" the way a colleague at the next desk does. It predicts the answer that best fits what is in the prompt, so anything you leave out, it fills in with the most common guess.
The AI guesses whatever you leave blank
When you type "write an apology email to a customer for late delivery", the AI does not know who the customer is, how late the order is, why, what the company is offering, or how formal to be. It fills in generic details and sometimes invents a promise ("100% refund") your company never made. Good prompt writers do not know magic words; they simply do not leave the AI guessing.
Not tied to any particular tool
Tools differ in interface and price, but the principles of prompt writing are almost the same everywhere. A good prompt is one that matches your actual job (Figure 1).
Photo: Bench Accounting benchaccounting / Wikimedia Commons, CC0
The 6-part formula: role – context – task – format – example – constraints
Not every prompt needs all 6 parts, but whichever part is missing, the AI will guess. For important work (anything going to a customer or your manager, or containing figures), write all six.
| Part | Answers the question | Short example |
|---|---|---|
| 1. Role | Who is the AI playing? | "You are a B2B customer care assistant." |
| 2. Context | What is happening, and who will read it? | "The customer ordered 20 chairs; delivery is 3 days late because stock reached the warehouse late." |
| 3. Task | Exactly what needs doing? | "Write an apology email proposing a new delivery date." |
| 4. Format | What should the result look like? | "Maximum 150 words, 3 paragraphs, polite form of address." |
| 5. Example | Which approved sample should it follow? | "Match the style of the sample email below." |
| 6. Constraints | What must it not do? | "Do not promise a discount; do not invent a delivery date." |
Role and context: put the AI in the right place
The role helps the AI choose its tone and level of expertise. Context is the part most often skipped but the most important: who will read it, what has happened and what the end goal is. Tip: write it the way you would brief a colleague who started today.
Task and format: one job, one output
Start the task with a specific verb: write, summarise, list, compare, turn into a table. The format states length, structure and tone. If you need to paste the result into Excel or Google Sheets, ask it to "return a table, one item per row, with no explanation".
Example and constraints: lock in the quality
One sample works better than ten sentences describing a style: paste an old approved email (with names and phone numbers removed) and ask it to "write in this style". Constraints are the fence: do not invent figures, no more than 1 page, and the most valuable line of all — "if information is missing, ask me first; do not guess". Put all six parts together and the apology email prompt looks like this:
Role: You are a B2B customer care assistant. Your style is polite and concise.
Context: My company sells office furniture. A customer (an accounting firm) ordered 20 swivel chairs
for delivery on 12 November, but delivery will be 3 days late because the shipment reached our
warehouse late. This is the customer's first order.
Task: Write an apology email and propose delivery on the morning of 15 November.
Format: Include a subject line; the body has 3 short paragraphs, maximum 150 words; formal address,
"we" for the company.
Style example: (paste an old approved email with names and phone numbers removed)
Constraints: Do not promise discounts or gifts; do not blame the supplier; do not invent information.
If any necessary information is missing, ask me before writing.5 before-and-after prompts for everyday office work
The table puts a one-line prompt side by side with a prompt that adds the missing parts. The originals are written in Vietnamese for Vietnamese offices; the same structure works in any language.
| Task | Bad prompt | Good prompt (shortened) | Parts added |
|---|---|---|---|
| "Write an email reminding a customer to pay." | "You are an accounts receivable clerk. A long-standing customer still owes the October invoice, 15 days overdue. Write a gentle reminder, 120 words, giving a new deadline of 20 November, with no threat of penalties." | Role, context, format, constraints | |
| Meeting minutes summary | "Summarise this meeting." | "From the minutes below, list: decisions made; a table of action items (task – owner – deadline); open questions. Do not add anything that is not in the minutes." | Format, constraints |
| Comparison table | "Compare 3 printer paper suppliers." | "From the 3 quotations pasted below (names hidden), build a table: criterion – Supplier A – Supplier B – Supplier C, covering unit price, delivery time, warranty and payment terms. If a quotation does not state something, write 'no information'; do not fill it in yourself." | Context, format, constraints |
| Excel formula | "Make an Excel formula for commission." | "Column C is sales, column D is commission: under 50 million VND earns 2%, 50 million or more earns 3%. Write the formula for cell D2, explain it, and give 2 examples so I can check it myself." | Context, format, a way to check |
| Customer service script | "Write a script for answering complaints." | "You are a phone customer care team leader. A customer calls to say a delivered chair has a loose leg. Write a 4-step script: greet – listen and confirm – solution (exchange or home repair within 3 days) – agree a time. Short, easy-to-say sentences; do not promise compensation beyond policy." | Role, context, format, constraints |
Email: say who reads it and what you want
Emails usually fail because of the wrong tone. State the relationship and the goal, then ask for "a shorter, more formal version" instead of editing sentence by sentence.
Meeting summaries: separate decisions from action items
"Summarise" is vague; the AI tends to rewrite the whole meeting slightly shorter. A meeting is worth three things: what was decided, who does what by when, and what is still stuck (Figure 2). Ask for exactly those three sections, using notes that contain no confidential information.
Photo: rawpixel.com rawpixel / Wikimedia Commons, CC0
Comparison tables and Excel formulas: AI drafts the frame, people own the numbers
With comparison tables, the AI tends to "fill every cell" with information it made up, so always add the constraint "if it is not stated, write 'not stated'". AI writes Excel formulas quite well but can mix up the separator (comma or semicolon depending on your regional settings) or the column, so always ask for "2 examples I can check". For quotations, let the AI draft the layout and terms; enter and calculate the numbers yourself.
Customer service scripts: write to be spoken, not read
A phone script needs short sentences, pauses for the customer to talk, and clear limits on what staff may promise. Read it aloud once; any sentence you stumble on, ask the AI to rewrite more simply.
Break the work down, make the AI ask questions, and request sources
The bigger the job, the more likely a single prompt gives a half-finished result. Turning it into a chain of steps and checking each one lets you catch mistakes early.
Split big jobs into 3–5 steps
For a monthly report: (1) send the data table with customer names hidden and ask for 5 highlights; (2) pick 3 and ask for an outline; (3) approve the outline, then have it write each section; (4) cut it down to 1 page.
Ask the AI to question you first
Add a line like the one below at the end of the prompt. The AI will point out what you forgot — the reader, length, deadline, level of formality — the fastest way to learn which parts you tend to leave out.
I need to draft an internal announcement about the Tet (Lunar New Year) holiday schedule for a 30-person company. Before writing, ask me up to 5 questions to clarify missing information. Only write after I have answered.
Ask for sources and separate "certain" from "guessed"
When asking about regulations or figures, add: "Give a source for each piece of information; mark anything uncertain as 'needs checking'". Even with a link, open and read it, because the AI may attach the wrong source.
Fact-checking: AI can make things up very fluently
An AI "hallucination" is when the model produces information that sounds plausible but is wrong: a legal article that does not exist, a statistic with no source, a miscalculation. Because the writing flows so well, mistakes are hard to spot if you skim.
Three kinds of information you must always check
- Figures and calculations: totals, percentages, differences — recalculate them yourself with a calculator or a spreadsheet formula.
- Citations and legal bases: names of legal documents, article numbers, studies — look them up on the official website of the issuing body.
- Names and dates: customer names, products, delivery dates, payment deadlines — compare against the original records.
A quick 2–5 minute check
Highlight every number, name and citation in the answer, then check them one by one. Anything you cannot verify in the time available, remove from the version you send.
Do not let the AI mark its own homework
Asking "are you sure?" can make the AI change a correct answer into a wrong one. Real verification means comparing against an outside source.
Security: data that never goes into a public AI tool
With the free or personal version of a public AI tool, what you send is stored on the provider's servers and, depending on settings, may be used to improve the model. Rule for beginners: treat anything you paste there as something outsiders might read.
Security warning: never paste into a public AI tool: passwords, OTP codes, API keys; personal data of customers and staff (full names with phone numbers, citizen ID numbers, addresses, bank account numbers, salaries, health records); contracts, quotations, cost prices, unpublished financial figures; confidential or internal documents. Customer data may only be processed on tools your company has approved, in line with internal rules and the personal data protection regulations in force in Vietnam. If unsure, ask IT or your manager before pasting.
Anonymise before pasting
Most work can still be done with substituted data: "Customer A", "Company B", rounded amounts. The AI needs structure and context to write; it rarely needs real names. Once the draft is done, put the real details back yourself in the copy on your own computer.
Be careful with attachments
Uploading a whole customer Excel file for "analysis" is the most common way data leaks, because files usually contain more columns than you notice. Keep only the columns you need and delete identifying columns. Payment data is also sensitive (Figure 3).
Photo: Shixart1985 / Wikimedia Commons, CC BY 2.0
Worked example: Thao saves more than 2 hours a week
Thao works in administration and sales at a 30-person furniture company in Da Nang. Each week she writes about 10 customer emails, takes minutes for 2 meetings, drafts 3 quotations and prepares 1 spreadsheet report. She timed herself for 2 weeks: the first week working as before, the second using the 6-part formula and her own prompt library (figures are an illustration; "after" times include re-reading and fact-checking).
| Weekly task | Times | Before (min each) | Total before | After (min each) | Total after |
|---|---|---|---|---|---|
| Customer emails | 10 | 8 | 80 | 4 | 40 |
| Meeting minutes | 2 | 30 | 60 | 15 | 30 |
| Quotations | 3 | 20 | 60 | 12 | 36 |
| Spreadsheet report | 1 | 90 | 90 | 50 | 50 |
| Total | 290 min | 156 min |
Making the sums add up
- Before: 80 + 60 + 60 + 90 = 290 minutes (4 hours 50 minutes).
- After: 40 + 30 + 36 + 50 = 156 minutes (2 hours 36 minutes).
- Saved each week: 290 − 156 = 134 minutes, about 2 hours 14 minutes.
- Over a 4-week month: 134 × 4 = 536 minutes, almost 9 hours — more than one working day.
One mistake she caught by checking
On one quotation she pasted the item table and let the AI "add up the total": 20 chairs × 450,000 VND and 2 meeting tables × 2,800,000 VND. The AI wrote a total of 14,400,000 VND. She recalculated: 20 × 450,000 = 9,000,000; 2 × 2,800,000 = 5,600,000; 9,000,000 + 5,600,000 = 14,600,000 VND. Had she sent the AI's version, the company would have lost 200,000 VND (about 8 USD) and some credibility. Since then her quotation prompt includes the constraint "leave price cells blank, do not calculate totals", and she never pastes real names or cost prices into a public tool.
A personal prompt library: write once, use all year
A prompt library is where you save prompts that gave good results, with blanks to fill in new details — a text file, a notes page or a spreadsheet, as long as you can find things fast (Figure 4) and it lives outside the AI tool, so changing tools does not cost you your work.
Photo: Shixart1985 / Wikimedia Commons, CC BY 2.0
Structure of a library template
Each template has: the task name, a prompt with blanks in square brackets, a note on "when to use" and "common errors". Example meeting minutes template:
[TEMPLATE] Internal meeting minutes
Role: You are the meeting secretary; write concisely and neutrally.
Context: Meeting on [topic] on [date], attended by [departments, no personal names].
Task: From the notes below, produce the minutes.
Format: 1) Decisions made; 2) Table: Task | Owner | Deadline; 3) Open issues.
Constraints: Add nothing that is not in the notes; mark anything unclear as "to be confirmed".
Notes: [paste anonymised notes]
Note: It often assigns the wrong owner; re-read that column.Which templates to build first, and how to update them
Start with the 5 templates in the table above. Every time you have to heavily edit a result, update the template too.
Common prompt-writing mistakes
A one-line prompt for important work
Sign: the answer is grammatically correct but generic and needs rewriting almost entirely. Cause: missing context and format. Fix: go through the 6 parts in turn, or add "ask me up to 5 questions before starting".
Trusting figures and pasting real data
Sign: documents go out with wrong numbers; prompts contain names, phone numbers and cost prices. Cause: trusting fluent writing, and convenience. Fix: check every number and citation; anonymise before pasting; use real data only on tools your company has approved.
Frequently asked questions
Is it better to write prompts in Vietnamese or English?
For Vietnamese office work, write in Vietnamese with full diacritics; popular tools understand Vietnamese well. If the result reads like machine translation, add the constraint "use natural Vietnamese and do not insert English words when a Vietnamese equivalent exists".
Do I need a paid plan to write good prompts?
No. The 6-part formula works with free versions; paid plan prices and features change over time, so check each tool's official website.
How can I tell when the AI is making things up?
There is no sure sign. Be suspicious of suspiciously round numbers, citations with specific article numbers and unfamiliar study names, and compare against the original records or official websites.
Practice exercises and self-assessment
- Pick the 3 tasks you do most this week and write a full 6-part prompt for each.
- Run both the one-line prompt and the 6-part prompt in the same tool; note how many minutes each result takes to fix.
- Take an answer that contains figures, highlight every number and citation, and check each one.
- Build a prompt library with at least 5 templates: name – prompt with blanks – when to use – common errors.
Self-assessment criteria
- Pass: 3 prompts with all 6 parts; at least 2 of the 3 tasks give a usable result after no more than one small round of edits.
- Pass: every figure has been recalculated or checked against a source; no real names, phone numbers or cost prices remain in prompts sent to public tools.
- Not yet: still sending one-line prompts for customer-facing work; trusting AI calculations without checking; no templates in the library yet.
Once you write prompts confidently, the next step on the roadmap is AI for office work: documents, spreadsheets, meetings — applying this formula to reports, spreadsheet analysis, note-taking and meeting summaries, with notes on copyright and citing sources.
Keep learning along the roadmap: from writing prompts to automating workflows with n8n, Make and Zapier and chatbots that answer from internal documents. See the 8-step AI-at-work specialist roadmap.
.jpg){kind=link}
.jpg){kind=link}
