Everyone’s talking about AI. Most of them are still just chatting with it.

I don’t mean that as an insult. I did the same thing for months. You open Claude or ChatGPT, ask it something, get a wall of generic text back, shrug, and close the tab. Somewhere in the back of your head a small voice says: this is supposed to be changing the world?

This site is part of me testing that, in public. I’m building a few things right now using AI as a working partner rather than a search box, and writing down what actually holds up.

Here’s what I found after actually using these tools to build something real, not just to ask them questions: the tool was never the problem. The way I was giving instructions was. And the fix is boring, unglamorous, and works almost every time.

Let me show you the difference, because it’s easier to see than to explain.

The same tool, two different results

A few weeks ago I needed help thinking through a business decision. Here’s roughly what I typed first:

“Give me advice on starting an affiliate marketing project.”

I got what you’d expect. Nine tidy bullet points. “Choose a niche.” “Create quality content.” “Be consistent.” All technically true. All completely useless, because I already knew them. It was the AI equivalent of asking a doctor “how do I stay healthy” and being told to eat well and sleep more. Thanks.

Then I changed one thing. Not the tool. The instruction:

“I’m a beginner with no audience. Before you give advice, research what actually works for beginners in this space right now, using real data. Then point out the flaws in my thinking, not just what I want to hear.”

Different universe. I’d assumed a beginner could publish a handful of posts, add some links, and see real money within a month. The AI didn’t just say “be patient.” It pulled actual industry data on how long first commissions typically take for beginners, then asked me a harder question: was I optimizing for income, or for building something worth trusting first? That reframed the whole project. I stopped asking “what should I sell” and started asking “what can I test in public until people trust my judgment.” That’s a very different 90 days.

Same model. Same five minutes. The only thing that changed was that the second instruction told the AI what job to actually do.

Why this works (the part nobody explains)

There’s a reason for this, and it’s not magic. Anthropic, the company behind Claude, puts it well in their own documentation. They compare prompting to giving instructions to a brilliant new hire who has zero context on your project. The clearer you are, the better they perform.

Sit with that comparison for a second, because it reframes everything.

If a genuinely smart new employee showed up on day one and you said “give me advice on our business,” they’d give you something vague and safe too. Not because they’re not capable. Because they have no idea what you actually need, what you already know, or what you’re really worried about. They’re guessing. And a smart person guessing still produces generic output.

The AI isn’t reading your mind. It’s filling in the enormous blanks you left with the most average, most probable answer. Vague question in, average answer out. That average answer is exactly what makes it feel like “AI slop.”

So the skill isn’t finding secret magic words. It’s learning to stop leaving blanks.

Three shifts that did the heavy lifting

I’m not going to hand you fifty prompts to copy. Copied prompts fall apart the moment your situation differs from the person who wrote them, and everybody’s peddling those anyway. I’d rather you understand the shifts, so you can build your own.

These three map directly onto what Anthropic actually recommends: be clear and direct, give examples, and give the model room to reason.

One: tell it who you are and what you actually need. “Explain email marketing” gets a textbook. “I run a small business, I’ve never sent a marketing email, explain it like you’re walking me through my first one” gets something I can use. The context isn’t padding. It’s the difference between an answer for everyone and an answer for you.

Two: ask it to disagree with you. This one is uncomfortable and that’s the point. By default these tools are agreeable. They’ll polish your bad idea until it shines. So I started ending instructions with “tell me where this is weak” or “argue the other side.” An AI that only nods is a mirror. An AI that pushes back is a second brain. You want the second one.

Three: make it show its work before it answers. For anything with moving parts, “think through the trade-offs step by step before giving your recommendation” beats “what should I do.” When it reasons in the open, you can see where it went wrong, and it makes fewer leaps. This isn’t a trick I invented. Giving the model space to reason is standard guidance, because a rushed answer and a reasoned answer come from the same model but land very differently.

The honest catch

I’ll level with you, because a post that only sells the upside isn’t worth your time.

None of this makes AI a genius button. It still gets things wrong with total confidence. It still needs you to know enough to catch the mistakes. The better your instructions get, the more it feels like a sharp collaborator, but it never stops being a collaborator you have to supervise. If you don’t understand the topic at all, better prompting just gets you more convincing wrong answers. That’s a real risk, not a footnote.

And the shift takes practice. My first tries at “better” instructions were clumsy and over-stuffed. You feel a bit silly writing three sentences of context for a question. Then the answer comes back actually useful, and you stop feeling silly.

Where to start

Pick one thing you’d normally just ask AI. Before you hit enter, add two lines: who you are, and what you actually want out of the answer. Then add “tell me what I’m missing.”

That’s it. One question, treated like you’re briefing a smart new hire instead of typing into a search box. Do that ten times and you’ll feel the gap between chatting with AI and working with it.

I’m still figuring a lot of this out myself, in public, as I build. When I find something that genuinely changes my results, I’ll write it down here. When something’s overhyped, I’ll say that too.

That’s the whole idea behind this place. I try it first.