DEV Community

Mirage Cloud IA
Mirage Cloud IA

Posted on

How to Write AI Prompts That Work for Everyday Business Tasks

Most advice about prompts treats them as a technical skill with secret formulas. In practice a prompt is a brief, and the people who get good results are the ones who would also write a good brief for a new colleague: say what the job is, give the background, show what good looks like and say how you will check it. A capable assistant with no context fills the gaps with guesses, and the guesses are where the poor results come from.

A six-part structure

  1. Context. Who you are, what the business does, who the output is for and which country and date the rules should come from.
  2. Task. One clear job, stated with a verb: write a reply, summarise, compare, list, check.
  3. Inputs. The material to work from, with a clear marker for where it starts and ends, so it is not confused with your instructions.
  4. Constraints. Length, tone, what to include and what to leave out.
  5. Format. Email, table, numbered list, two paragraphs. Say which.
  6. Check. Ask it to list its assumptions and to flag anything it could not establish from the material.

You will not need all six every time. For a quick rewrite two lines are enough. For anything with a number, a date or a legal point in it, use all six.
Worked example: a follow-up email
Weak prompt: Write a follow-up email for a quote.
Better prompt: I run a three-person landscaping company in Lyon. Write a follow-up email to a homeowner who received a quote for a garden redesign five days ago and has not replied. The quote was 6,400 euros and includes a free site visit before work starts. Tone: friendly and brief, no pressure. Under 90 words. End with one question that can be answered yes or no. Do not invent discounts or deadlines.
The second version gives the assistant a business, a reader, a set of facts, a tone, a length and a limit on invention. The result needs a light edit rather than a rewrite.
Worked example: a contract
Prompt: Below is a supplier agreement. List the obligations that fall on us, with the clause number for each. Then list every date, deadline and notice period. Then list penalties and termination conditions. If something is unclear, say so instead of guessing. Text: [paste].
Notice what this asks for: extraction with references, not an opinion. You can check every line against the document, which is the point. Then take the list to someone qualified before you rely on it.
Worked example: commentary on figures
Prompt: Here are this month's figures: [paste]. Write three sentences for the owner explaining what changed against last month. Use only the numbers provided. If a reason is not in the data, say it cannot be determined. Do not guess causes.
The last three sentences are what stop a general assistant from inventing explanations.
Habits that make results better
Name the jurisdiction and the date. For rules, say French law and the current year. It reduces answers drawn from another country or an older version.
Show an example of your own good work. One paragraph written the way you want it is worth a page of description.
Work in two or three turns. Ask for a plan or options first, choose one, then ask for the full text.
Ask for alternatives. Three subject lines are more useful than one.
Keep a prompt library. When a prompt works, save it with its date and what it was for.
Ask for sources, then check them. Verify any figure, date or reference before using it.
Mistakes to avoid
Bundling five jobs into one request.
Burying the key fact in the middle of a long paste.
Setting contradictory constraints, such as very short and very detailed.
Pasting passwords, personal data you have no basis to process or other people's confidential material.
Expecting the assistant to remember last week's conversation. State the context again or use a tool that keeps it.
Treating confident output as correct output.
When the tool already knows your business
Much of the effort in a prompt goes on re-explaining context. Tools that are connected to your accounting, banking or calendar remove that step, and tools built around specific roles carry their scope with them. Mirage Cloud works that way, with agents scoped by function and connections to tools such as Qonto, Pennylane and Google Calendar. The briefing discipline above still helps, but you spend less of it on background and more on the task.

Top comments (0)