How to write better prompts for repeatable results at work

On-Off Group

September 2026

Write the prompt once as a reusable instruction sheet: the task, the real input, the rules and an example of the output you want. Then test it on five real cases from last week and fix the wording where it fails.

If you use AI at work and only get a good answer about half the time, this is about fixing that. You can write a prompt. What you cannot do yet is get the same quality twice, so you still check every line by hand and the time saving never shows up.

The fix is simple. Write the prompt down once, the way you would brief a new joiner on a task they have never done. Then test that briefing on real work before you trust it.

This works for the repeat tasks in operations: summarising a complaint, drafting a reply to a client, checking a form against a checklist.

Retyping the prompt every time is why the answers keep changing

A complaint email lands. You open the AI tool, type something like "summarise this and suggest a reply, make it polite", paste the email, and get something usable. The next complaint comes in later and you type it again, slightly differently. This time the reply invents a resolution date nobody promised.

The model did not change. The instruction did. One day you said "polite", the next you said "empathetic", and you never said anything about dates at all.

So write the prompt down once, in a file, with a name. In our AI prompting sessions with bank and insurer teams, the people still using AI on Monday are the ones who saved their prompt in a file. The rest went back to typing it fresh each time. Keep it where the rest of the process lives, so nobody is improvising while a customer waits.

Write the prompt out like a briefing

Write it in four parts instead of one line. Use four:

The task. One sentence. "Summarise this customer complaint and draft a reply for a branch officer to send."
The input. Paste the real thing. The email, the form, the notes. Do not describe it.
The rules. What must be true of the answer. Maximum 120 words. No commitments on timing. Use the customer's name once. Flag anything that looks like a fraud claim instead of replying.
The output shape. "Give me two parts: a three line summary, then the draft reply."

That is usually ten to twenty lines. It feels like a lot of typing the first time. After that you are only pasting it.

One tip on the rules: most of them come from mistakes. Every time the AI does something you had to correct, add a line banning it. The prompt ends up holding all the corrections you made, so you stop making them.

Show it one good answer

Adding one good example usually helps more than adding another rule.

Find a reply you or a colleague wrote that you would be happy to send again. Paste it under the heading "Example of a good reply". Then add a line: "Match the tone and length of the example."

Nobody can write down what the right tone is, and every officer reads "professional but warm" differently. A real reply from your own inbox settles it. A real reply from your own inbox does not.

If the task has a format, show that too. A handover note with the right headings. A checklist result written the way your team writes it. One example is enough. Two is better if the cases differ a lot, such as a simple query against a disputed charge.

Run the prompt on five real cases from last week

Most prompts get tested once, on the case that happened to be open. Then they go into use and fail quietly on the awkward ones.

Take five real items from last week. Pick them on purpose: two straightforward, two messy, one you would escalate. Run the prompt on each and mark the answer usable or not usable. Write down why it failed, in your words. "Too long." "Made up a turnaround time." "Missed that the customer had already called twice."

Then change one thing in the prompt and run the same five again. Same cases, before and after, so you can see whether your edit helped or you just liked it more.

The same habit is what makes bigger project work provable, and it scales up. In a series of five day design sprints with a large insurer, journeys that used to take six months or more were designed, tested with customers and live in four weeks, and completion rates on the new journeys rose 80 per cent. Those numbers exist because someone wrote down the old ones first. Five cases marked usable or not usable is the same idea, at prompt size.

Decide what the AI is not allowed to do

You also need to know what it does when the input is bad. A form arrives with the policy number missing. A complaint mentions a possible fraud. A client asks for a refund amount.

Put those cases in the prompt with an instruction each time. "If the policy number is missing, do not guess. Return: MISSING POLICY NUMBER and stop." "If the message mentions fraud, do not draft a reply. Return: ESCALATE, with a one line reason."

Then you can scan the output. In a page of normal looking drafts you will miss the one that should never have been sent. ESCALATE in capitals does not hide.

Also decide what stays out of the tool. If there are rules about customer data in outside tools, put them at the top of the prompt file so the next person sees them before pasting.

What to do this week

Pick the one task you already use AI for most often, and only that one.
Write the prompt out properly: task, input, rules, output shape. Save it in a file with a name your team would find.
Paste in one real answer from your own work as the example of good.
Run it on five real cases from last week, two easy, two messy, one to escalate, and mark each answer usable or not.
Add one rule for every correction you had to make, then run the same five again. More on how we run the AI prompting sessions is at onoffgroup.com.

Questions people ask

Why does the same prompt give different answers?
Usually because the prompt is retyped from memory each time, so the wording and the rules change. A saved prompt with the task, rules and a sample output gives much steadier results.

How long should a work prompt be?
As long as the briefing you would give a new joiner on their first week. That is often ten to twenty lines, including one example of a good answer.

How do I know a prompt is good enough to use?
Run it on five real cases from last week and mark each answer usable or not. If four or five are usable with light editing, start using it and keep a note of the failures.

On-Off Group trains teams, tests products with real customers, finds where a transformation has stalled and builds what gets it moving, for banks, insurers and enterprises in the Philippines, since 2015. How we help with ai prompting.