Writing good prompts: a practical guide for business teams

A good prompt is a clear brief. It tells the AI who it is helping, what the task is, what information to use, what the result should look like and what to avoid, in the same way you would brief a capable new staff member on their first day. Most disappointing AI answers come from short, vague prompts rather than from a weak model. Prompt writing is a skill your whole team can learn, and once a prompt works well, it can be saved and reused by everyone.

The six parts of a good prompt

You will not need all six every time, but checking for them fixes most weak prompts.

1. Context. Who you are and who the reader is. "We are a home appliance shop in Rangpur. Most customers are families who pay cash on delivery or in instalments."
2. Task. One clear job, starting with a verb: draft, summarise, list, translate, compare.
3. Input. The material to work with, clearly separated from your instructions, for example between quotation marks or under a heading.
4. Rules. What it must and must not do. "Use only the prices given below. If the answer is not in the information, say a team member will confirm."
5. Format. Language, length, tone and structure. "Reply in Bangla, in a short paragraph, warm and polite."
6. Example. One sample of a good answer, if you have one. Examples often teach tone better than any description.

A before-and-after example

A weak prompt:

Write a reply to this customer.

A stronger version of the same request:

You are helping the customer service team of a home appliance shop in Rangpur.
Task: draft a reply to the customer message below.
Use only the facts in the "Shop information" section.
If something is not covered there, say that a team member will confirm.
Do not promise delivery dates, discounts or warranty terms.
Tone: polite, warm and brief.
Language: reply in the same language the customer used.

Customer message:
"..."

Shop information:
"..."

The second prompt takes a minute longer to write, but it is reusable, and it prevents the most common problems: invented prices, promises nobody approved, and replies in the wrong tone.

Tips for Bangladeshi teams

• Name the language and script. "Bangla script", "English" and "Banglish in Latin letters" are different requests. Decide which one your business uses for each channel.
• Set the level of formality. Tell the AI to address customers with the respectful form of "you", and to avoid stiff, bookish Bangla in chat replies.
• Give local terms. Product names, district and upazila names, and how you write payment methods such as bKash, Nagad and cash on delivery.
• Test with real messages, including spelling mistakes and mixed languages, before relying on a prompt.

For product listings in two languages, the Bangla and English product content solution shows how prompts fit into a write-and-review routine.

Make uncertainty visible

AI models tend to answer even when they should not. Good prompts make it easy for the model to admit gaps:

• Tell it exactly what to do when information is missing.
• Ask it to list any assumptions it made.
• Ask for answers in fixed sections or fields, so missing parts stand out.
• For facts, ask it to quote the line of your document it relied on.

The article on why AI sometimes states wrong things confidently explains why this matters.

Build a shared prompt library

Prompts that work should not live in one person's chat history.

• Keep them in one shared document or sheet, each with a name, purpose, owner and the date it was last tested.
• Use placeholders such as [customer name] or [product] so staff know what to fill in.
• Store a sample input and a good output next to each prompt. When someone edits the prompt, they rerun the sample and compare.
• Use the library for training. New staff start with tested prompts instead of inventing their own.
• Keep sensitive data out. Samples should use made-up or anonymised details.

Review what comes back

A prompt is not finished until someone checks the output. Before anything goes to a customer, ask four questions: are the facts correct, did it follow the rules, is the tone right, and would I send this under my own name? The solution on AI-written email drafts that staff approve shows how to build that check into daily work.

A quick checklist before you press enter

• Have I given the context: who we are and who will read this?
• Is there one clear task, starting with a verb?
• Is the input clearly separated from the instructions?
• Have I stated the rules, including what to do when unsure?
• Have I specified language, length, tone and format?
• Would an example help?

If your team wants to move from individual prompts to a reliable automated workflow, see AI and workflow automation.