How to write instructions an AI actually follows

The gap between people who find AI useful and people who find it disappointing is mostly not the tool. It's how they ask.

This isn't about magic phrases or "prompt hacks." It's closer to how you'd brief a competent new employee who is fast, well-read, has no context about your business, and will confidently do the wrong thing rather than ask.

Say who it's for and what happens next

The single most common mistake is asking for a thing without saying where it's going.

"Write a description of my business" produces something generic. "Write a description of my business for the front page of my website, aimed at someone who has never heard of us and is comparing three options" produces something usable, because now it knows the reader, the placement, and the job the words have to do.

Audience and destination. Almost every weak result is missing one of them.

Give it a real example

If you want something in your voice, show it your voice. Paste in two emails you've actually sent. Paste in the last quote you wrote. Say "match this tone."

Describing a tone in adjectives, like "friendly but professional," means almost nothing, because those words mean different things to everyone. An actual sample of your writing means exactly one thing.

This applies to structure too. If you want a quote laid out a particular way, show it the last quote you sent rather than describing the layout.

Tell it what you don't want

Constraints do more work than instructions. "Don't use bullet points." "Don't mention price." "Keep it under 120 words." "Never promise a timeline."

Negative instructions are unusually effective because they close off the specific ways things go wrong. If you got a bad result, the fastest fix is usually to name what was bad about it and ask again, not to rewrite the whole request.

Ask for the reasoning when it matters

For anything where being wrong is costly, ask it to explain how it got there before it gives you the answer. "Walk me through your thinking, then give me the recommendation."

Two things happen. The output gets better, because working through it in order produces fewer leaps. And you can actually check it. You can see the assumption it made about your business that isn't true, instead of just receiving a confident conclusion built on it.

Work in passes, not in one shot

Nobody writes a good brief on the first try, including people who do this professionally.

Get a rough version. Say what's wrong with it. Get another. Three rounds of "closer, but make the opening less formal and cut the last paragraph" beats twenty minutes crafting the perfect opening request.

The corollary: don't start over when something is 80% right. Fix the 20%.

Give it the facts it can't know

AI doesn't know your prices, your service area, your capacity, your busy season, or that you stopped offering that thing in March. If those matter to the answer, put them in the request.

A useful habit: keep a plain text file with your business basics: what you do, who for, where, prices, anything you always have to explain. Paste it in at the start of anything important. It takes ten seconds and removes most of the wrong assumptions.

Say how long it should be

"Short" is not a length. "Three sentences" is. "One paragraph." "Under 200 words." "A single line I can put in a text message."

Left unspecified, you'll get medium-length, every time, regardless of whether that fits.

A worked example

Weak:

Write a follow-up email to a customer.

Better:

Write a follow-up email to a customer who asked for a quote last Tuesday and hasn't replied. I sent them a price for a bathroom refit. Friendly, not pushy, three sentences maximum, no discount offered. Here's how I usually write to customers: [paste two of your emails]. Don't use the phrase "just checking in."

The second one takes forty seconds longer to write and will save you three rounds of editing.

The mindset that helps most

Treat a bad result as a briefing problem before you treat it as a tool problem. Nine times out of ten, something you knew and didn't say is the reason.