You're the expert it needs

A couple of years ago, getting good results from AI seemed to call for special tricks: magic phrases, elaborate role-play, secret formats. The tools have got much better at understanding plain requests, and the tricks matter less every year. What hasn't changed is that the AI doesn't know your work. It doesn't know what good looks like in your job, which mistakes would be a disaster and which would be a shrug, or when it should stop and ask.

You do. That knowledge, written down, is now the thing that makes the biggest difference. You don't need to become technical. You need to say out loud what you already know.

Four things the AI can't know about your work

• Your situation: your manager, your clients, your customers, your market, your town.
• Who pushes back: the customer who reads every line, the regulator, the competitor, the boss who hates jargon.
• What you've already promised: the price you quoted, the deadline you agreed, the policy you follow.
• Your definition of good: what a great result looks like here, not in general.

The AI is very good at the familiar patterns: a polite email, a standard quote, a clear summary. It will fill in those four things with sensible-sounding guesses. The guesses are where your results go generic or wrong.

The expert brief

Write this once for each kind of task you hand to AI often, and paste it in (or save it in your assistant's instructions). It takes ten minutes and pays back every time.

I'm a [your role] and this is how I want you to help with [kind of task].

Good looks like: [what a great result has, specifically, e.g. "answers the question in the first line, under 150 words, one clear next step"]

The worst mistakes would be: [the two or three errors that would really cost you, e.g. "promising a delivery date", "getting a dose or a dollar figure wrong"]

Mistakes that are merely annoying: [e.g. "slightly too formal"]

What you can't know about my situation: [your clients, the rules you work under, what you've already promised]

Stop and ask me if: [e.g. "the customer mentions a refund", "you'd need a number I haven't given you"]

Never: [e.g. "invent a policy", "use the client's first name"]

Three people, three briefs

Three jobs that stay yours

• Own what you can't take back. The AI can draft the email; you decide when it's sent. See One-way doors.
• Define success clearly enough to check. "Better" can't be checked; "under 150 words with a date" can. See Turn "make it better" into a check.
• Keep testing your own definitions. When a result passes your rules but still feels wrong, the rules need work.

Common mistakes

• Hunting for a better magic phrase when the missing piece is what you know.
• Describing quality ("professional", "engaging") instead of behaviour ("no exclamation marks, one question at the end").
• Leaving out the worst mistakes because they seem obvious. They are obvious to you.
• Writing the brief once and never updating it after the AI gets something wrong.