Train AI with your own examples, not better wording

If AI is not producing work that looks like your business's work, the fix is almost never a cleverer sentence. It is showing it three real examples of the finished thing, including one that was not straightforward. Examples carry the judgment that instructions cannot describe, and unlike a clever prompt they keep working when the model changes underneath them.

Is there a magic prompt wording?

No. And the belief that there is costs small businesses real money.

A mate asked me last month for the magic phrasing that makes AI write his quotes properly. He was convinced there was a form of words he did not know, and that someone was selling a course on it.

There isn't one, and the actual answer is more boring and more durable. The people who build these models say so themselves. Anthropic's own prompt engineering guidance ranks its techniques roughly in order of how broadly they help, and using examples, what they call multishot prompting, sits second, behind only being clear and direct. Their guidance is to include three to five diverse, relevant examples, and to make sure they are diverse enough to cover edge cases.

Read that again with a small business hat on. Three to five. Relevant to your actual case. Covering the awkward ones. That is not a technical skill. That is a filing job.

Why do examples work better than instructions?

Because most of what makes your work yours cannot be written as a rule.

You could try. Be professional but not stiff. Mention the site access issue if there is one. Do not commit to a date until we have seen it. Match the tone we use with repeat clients but not new ones. By the time you have written the instruction that captures your judgment, you have spent longer than it would take to do the job, and you still will not have caught the case you did not think of.

An example carries all of that implicitly and costs you nothing to produce, because you already made it. It is the difference between describing a house and showing someone a photo.

The onboarding comparison is the one that lands with most owners. You would never hand a new starter a beautifully worded instruction and hope. You would show them three jobs you have already done, including the one that got messy, and tell them what finished work looks like around here. Everyone already knows this. Very few people have connected it to the AI sitting on their desk.

Why does the messy example matter most?

Because the tidy ones teach nothing.

If all three of your examples are straightforward jobs, you have taught the AI to handle straightforward jobs, which it could already do. The value is in the awkward case: the quote where site access was difficult, the email to the client who always asks the same follow-up question, the job that had to be re-scoped halfway through.

Those are the moments where your business made a judgment call, and they are the only place that judgment is visible. This is also why one example is not enough. Research on few-shot prompting shows the gains grow as you add examples, and that a single example occasionally makes output worse rather than better, because the model over-fits to whatever was peculiar about that one case. Three is roughly where it starts behaving.

What should I actually collect?

Start with the task you repeat most, and gather the artefacts you already have.

1. Pick one repeated job. Quotes, enquiry replies, job summaries, scope documents. One, not five.
2. Find three finished examples you were happy with. Real ones out of your own files, not idealised versions written for the occasion.
3. Make sure one of them was awkward. If all three were easy, go and find a fourth that was not.
4. Add one line of context to each. What the situation was and why you handled it the way you did. This is the part that turns a document into a lesson.
5. Paste them in before you ask for anything. Then compare with what you were getting yesterday.

That is the whole method. It takes about an hour and it is the closest thing to a durable AI asset a small business can build, because it survives model upgrades, tool changes and staff turnover. Most of what we do at Handiwork ends up being some version of this rather than anything clever.

Once you have a pile of examples for one job, you are most of the way to turning that expertise into a reusable AI skill, which is the same idea made permanent instead of pasted in each time. It also pairs with giving AI the standing context about your business: context tells it who you are, examples tell it what good looks like. You need both, and neither is a wording problem.

Does it need to be done properly first?

No, and waiting until it is tidy is the most common way this never happens.

Three examples beats none by a wide margin. You will find out what the fourth should be by watching where the first three fall short, which is a much better way to design the set than planning it in advance. If you are not sure which job to start with, the same instinct applies as in choosing where to start with AI at all: pick the thing that annoys you weekly, not the thing that sounds impressive.

You can see the shape of what this leads to on our use cases page.

Ready to find out where you stand?

If you want a straight read on which of your repeated jobs is worth writing down first, our free AI Readiness Check takes about five minutes. No cost, no pitch.