Better instructions, so your AI gets it right more often

What better instructions are

Every AI helper runs on a set of written instructions. They tell it what job it’s doing, what to read, how to write, and when to stop and ask a person. Better instructions (the industry calls it prompt optimisation) is the work of writing those well, testing them on real examples, and fixing them when the AI gets something wrong.

It sounds small, and on paper it is. A good set of instructions for one AI helper is often a page or two. But that page decides whether the drafts in your inbox each morning are ready to send or need rewriting from scratch.

Where small teams get stuck

You’ve probably had this experience with ChatGPT or Claude. The first answer is close, so you add a line, then another, then you paste in an example. It gets better, and then a week later it does something odd and you’re not sure which line caused it.

That’s fine for one-off questions. It’s a problem when an AI helper is answering customer emails or drafting client updates every day, because the same gap in the instructions repeats every time. We work with a three-person online skincare store where the first version of its AI helper sent simple tracking replies on its own. Then a customer wrote in about a jar that had arrived broken and got a cheerful tracking link back. The instructions hadn’t told it how to spot a complaint.

Why adding more words doesn’t fix it

The instinct is to keep adding rules. But long instructions get slower, cost more to run, and start to contradict themselves. Sometimes the fix is to take something away.

Saltbush Studio, a six-person design agency in Fremantle, has an AI helper that drafts its Friday client updates from the project board. We also fed it the team’s Slack chat, thinking more context would help. It picked up jokes and half-made decisions and put them in front of clients. So Slack came out, and the project board is now the only source. The drafts got better with less to read.

How we tune them

We treat instructions the way a good team treats a recipe. They’re written down, tested, and changed on purpose.

1. Start with one job. We write instructions for one AI helper doing one task, like replying to order emails or drafting product descriptions.
2. Collect real examples. We gather past emails, listings, or reports, including the awkward ones, so we can test on your real work.
3. Write and test. We run the AI helper on those examples, check where it’s right and wrong, and adjust.
4. Mark the hand-offs. We write down what the AI must pass to you, like a complaint, a refund, or a question about allergies.
5. Keep a record. Every change to the instructions is saved with a note on why, so if a change makes things worse, we can roll it back.

What changes in your week

For Saltbush Studio, the twelve Friday updates now take the director under an hour, and the director still reads each one and adds a personal line. For the skincare store, every reply now waits for the founder to read and send, and the founder clears the order emails in about twenty minutes each morning. Returns still go through by hand, because each one needs a judgement call about the product.

Better instructions won’t make any AI perfect. They make it wrong less often, and they make sure the jobs it can’t handle reach a person.

Better instructions by industry

Every industry has its own words, its own tone, and its own questions a person must answer. See how it applies in fashion, real estate, marketing and creative agencies, e-commerce, and beauty and wellness.