Treat it like a person, not a machine
What this shows
Chapter 1 makes a claim that’s hard to believe until you see it: the AI you use today and the AI you use after fifteen minutes of setup are, in practice, two different tools. Same model. Same prompt. Different output.
You’re about to watch that happen. The same email, the same review request, run through the same AI twice. First with no setup. Then with the kind of setup any reader of this book can paste in once. The gap between the two responses is the whole argument of this chapter.
The demonstration
The email
Imagine you’ve drafted this email and you’re about to send it. You’d like a second opinion first.
Copy this email
Subject: Pilot extension — request for another quarter
Hi Sarah,
I wanted to follow up on our pilot programme for the new customer-onboarding system. As you know, the results have been really positive — we’ve seen significant engagement from the pilot group and the feedback has been overwhelmingly good.
I’d love to extend the pilot for another quarter so we can build on this momentum. The team is excited about where this could go, and I think with a bit more time we can really demonstrate the value to the wider business.
Would you have any time in the next week or two to discuss next steps?
Thanks, [Your name]
So you open your AI tool and ask for a review.
Take 1 — stock AI
Can you review this email and let me know what you think? I’d like to send it later today.
Here’s what you get back:
AI: Great email! It’s professional, polite, and clearly communicates your request. The tone is positive and collaborative, which is perfect for a follow-up message to a senior colleague.
Here are a few suggestions to make it even stronger:
Quantify the results. Phrases like “really positive” and “overwhelmingly good” could be more impactful with specific numbers — for example, engagement rates or completion metrics.
Be specific about next steps. Rather than “discuss next steps,” you could propose a concrete outcome for the meeting, such as agreeing on success criteria for the next quarter.
Reiterate the business value. A brief sentence on the broader strategic benefit would help reinforce why this matters beyond your immediate team.
Overall, this is a solid draft — just a few tweaks and you’re good to go!
You read it. It sounds sensible. Maybe you’ll add a metric and tighten the closing. Off it goes.
But notice what you didn’t get. You didn’t get told whether the email actually does its job. You didn’t get pushed on anything. You didn’t get warned about anything. You got a polished version of yes, that’s fine.
The setup
Now imagine you’d spent fifteen minutes a few weeks ago putting this into your AI’s custom instructions:
Keep responses short and direct. Don’t open with “Here is…” or close with a summary paragraph. Use British English.
When I propose a plan, a decision, or a piece of writing, lead with what’s weakest about it before you tell me what works. Separate facts from judgements and assumptions, give your reasoning for judgements, and flag where you’re uncertain. If you critique a draft, always offer specific rewritten options that fix the issues — don’t just diagnose. For casual or quick requests, skip this and just answer.
Never use analogies unless I ask for one. Never write “in conclusion” or “in summary” — just stop.
That’s it. Same AI. Different setup.
Take 2 — configured AI
You send the same email and ask the same question:
Can you review this email and let me know what you think? I’d like to send it later today.
Here’s what comes back now:
AI: This draft has two main weaknesses before it has any strengths.
First, the opening rests on a claim — “the results have been really positive” — that the email never substantiates. You write “significant engagement” and “overwhelmingly good feedback” without naming a number, a behaviour, or a quote. If Sarah has any reason to doubt this is worth her time, she’ll close the email at that paragraph. You’re asking her to extend a budget on someone else’s say-so.
Second, the ask. “Would you have any time in the next week or two to discuss next steps?” isn’t a request — it’s a hedge. It puts the cost of action on her diary and gives her nothing concrete to react to. If she’s busy (she is), the email goes to the bottom of the pile.
Two rewrites depending on what you have:
If you have a specific result you can cite: Open with one number. “Pilot participants completed 73% of onboarding modules vs 41% in the control group.” Then make the ask explicit: “I’d like to extend for Q2 with the same group plus 30 new users. Could we lock this in at our 1
on Thursday?”If the data isn’t ready yet: Acknowledge that and act anyway. “The full metrics will be ready next week but the early signal is strong enough that I want to plan for an extension rather than a wind-down. Could you give me twenty minutes on Thursday to walk through what I have so far?”
One thing I’m uncertain about: I don’t know what Sarah’s relationship to this project is. If she championed it originally, the soft version is fine. If she’s sceptical or new to it, you need the direct version with a number.
You read it. You sit with it. The first response was easier to ignore. The second one is harder to ignore because it’s right.
The point
That’s the same AI. Same model. Same prompt. Same email.
The only difference is that one of them had been told, once, what kind of help you actually want. Don’t agree by default. Push back. Tell me what’s wrong before you tell me what’s working. Offer specific rewrites, not just diagnosis. Skip the preamble. Use British English.
Fifteen minutes once. Every conversation afterwards starts from a better place. Chapter 1 made the claim. Take 2 is what the claim looks like in practice.
Try it yourself
The setup that produced Take 2 lives in the Setting up your AI category of the Prompt Library. The version shown above blends three of those starter prompts — voice and tone, challenge me, and what never to do — into a single block.
Copy the prompts you want, paste them into your AI tool’s settings menu, save, and then run something real through it. An email you’re drafting, a plan you’re sketching, a decision you’re working through.
The first time AI pushes back on you instead of agreeing with you, you’ll know it’s working. It can be slightly disconcerting — most AI you’ve used to this point has been trained to make you feel good. This one is being trained, by you, to make you think clearly. That’s the difference.