Stop Yelling Instructions at a Confident Idiot

Most people prompt AI like a Magic 8-Ball. Here's the 5-minute system that actually gets it to follow your instructions.

Most AI advice tells you to write better prompts.

“Be specific.” “Add context.” “Use the magic formula.”

But here’s what nobody mentions: you’re asking AI to make 400 tiny decisions without consulting you.

What tone? It guesses. What length? It guesses. What your audience already knows? It guesses. What would make you delete everything and start over? It guesses spectacularly wrong.

The result: polished, fluent, grammatically pristine output that solves the wrong problem. Then you spend 30 minutes rewriting 80% of it while wondering if starting from scratch would’ve been faster.

The solution isn’t another prompt template or “Ultimate AI Pack” for $97. It’s embarrassingly simple: make AI ask questions before it does anything.

Today’s guest post is from Nick Quick, who writes Co-Write with AI - a newsletter teaching content creators how to collaborate with AI without producing generic slop. His Voiceprint methodology helps writers document their voice so AI can actually follow it.

What I like about Nick’s content is his anti‑hype approach to using AI. It’s not about saving “10 prompts to grow your audience” or “50 prompts to conquer 2026.” It’s about learning how to collaborate with AI.

Check out his latest three posts if you are curious to see what he writes about:

1. If ChatGPT Agrees With You, Rewrite It

2. One Sample. One Prompt. Twenty Minutes.

3. Your AI Writing Isn’t Special. Here’s How to Fix That.

Nick’s built a three-step system that transforms AI from a confident guesser into an actual thinking partner. No fancy tools. No complicated setup. Just Feed, Reflect, Correct.

Here’s Nick.

---

Hello, Nick here 👋

I’m going to describe a management style. Tell me if you recognize it.

Day one. New hire. Brilliant resume. Glowing recommendations. You walk them to their desk and say: “Write me a marketing email.”

Then you leave.

No context about your company. No examples of what’s worked before. No explanation of who you’re talking to, what you sound like, or what happens if this goes wrong. Just vibes and a deadline.

They produce something. It’s polished. Professional. Uses words like “synergy” unironically.

It’s also completely wrong.

This is how most people use AI.

(Narrator: It is not going well for most people.)

We treat AI like a Magic 8-Ball. Shake it, ask our question, stare into the little window hoping the plastic triangle floating in blue goo says something useful. Reply hazy, try again. The 8-Ball approach gets you 8-Ball results: random, generic, occasionally correct by pure statistical accident. You’re not collaborating. You’re just gambling on probability and calling it a workflow.

But here’s the thing nobody mentions in those “10X Your Productivity” threads: AI isn’t a mystical orb. It’s more like a junior collaborator with a photographic memory, zero context about your life, and the confidence of a guy who just discovered cryptocurrency.

Capable. Eager. Completely clueless about what you actually need.

And we keep shaking it harder like that’s going to help.

---

The Fundamental Problem (It’s Us, Actually)

When you give AI a prompt and hit enter, you’re asking it to make approximately four hundred tiny decisions without consulting you.

What tone? (It guesses.)

What length? (It guesses.)

What matters most? (It guesses.)

What your audience already knows? (It guesses.)

What would make you throw your laptop out a window? (It guesses wrong.)

AI is remarkably good at guessing. That’s the problem. It guesses with the confidence of someone who has never experienced consequences. It produces polished, fluent, grammatically pristine garbage that looks like it knows what it’s doing.

The output arrives fully dressed for a job interview. But it prepared for the wrong job. At the wrong company. In the wrong industry. And it’s somehow still smiling.

So you fix it. And fix it again. And eventually you’re rewriting 80% while wondering if you could’ve just written the damn thing from scratch in less time.

(You could have. You definitely could have. But now we’re here.)

The solution isn’t better prompts. It’s not a magic formula. It’s not paying $97 for someone’s “Ultimate Prompt Pack” that promises to revolutionize your workflow.

(Those packs are just someone else’s guesses packaged as certainty. Slop gift-wrapped with a little bow on top.)

The solution is embarrassingly simple: make AI ask questions before it does anything.

---

The Onboarding AI Should Have Gotten (But Didn’t)

When you hire an actual human (a practice I’m told still happens occasionally), you don’t just chuck tasks at them and hope for the best.

You have intake conversations.

“Here’s what we’re trying to accomplish.”

“Here’s who we’re trying to reach.”

“Here’s what ‘good’ looks like, and what ‘catastrophic failure’ looks like.”

“What questions do you have?”

That last one matters most. Good hires ask clarifying questions. They surface what they don’t know before they start confidently producing wrong things. They recognize that their brilliant ideas mean nothing if they’re solving the wrong problem.

AI can do this too. It’s actually quite good at it.

It just doesn’t do it by default. By default, it skips the intake meeting and sprints directly to execution. It fills every gap with its best statistical guess (which is, by definition, the average of everything it’s seen). Your output sounds like... everything else. Because it literally is everything else, averaged together.

The fix: Force the intake conversation. Make AI ask before it answers.

Three steps. Five minutes. Here’s the system.

---

The Full System: Feed, Reflect, Correct

Step 1: FEED

Give AI your task context AND an instruction to ask questions first.

[Your task here - what you're creating, why it matters, any context you already know]

Before generating anything, ask me 3-5 clarifying questions about:
- Who exactly will read this and what they already know
- What outcome I actually want (not what you assume I want)
- Constraints I haven't mentioned but will definitely be mad about later
- What "success" looks like vs. "this made everything worse"

Don't produce anything until I've answered. Your questions should help you understand my intent, not demonstrate your knowledge.

That last line is critical. Without it, AI defaults to asking questions that show off what it knows rather than surfacing what you know.

Bad AI questions (flexing, not learning):

• “Would you like me to use the AIDA framework or the PAS structure?” (Wow, you’ve heard of frameworks. Incredible. Not helpful.)

• “Should I include statistics to support the argument?” (It’s guessing what might seem professional.)

• “How many paragraphs would you prefer?” (It’s asking you to do its job.)

Good AI questions (actually trying to understand):

• “Who’s reading this, and what do they believe about this topic right now?”

• “What’s the one thing that absolutely has to land?”

• “What would make you look at this output and immediately delete it?”

• “Are there phrases or approaches you hate and will spend 20 minutes removing if I include them?”

When AI asks good questions, something shifts. It stops performing “capable assistant” and starts acting like someone who actually wants to help.

(The bar is low. And yet here we are, celebrating when software asks us what we want.)

Answer the questions. Be specific. If AI asks “who’s the audience?” don’t say “professionals.” Say “mid-level marketing managers who are skeptical of AI but under pressure to use it, probably while their CMO breathes down their neck about ‘innovation.’”

Specificity is everything. Vague inputs produce vague outputs. This is not complicated, and yet.

Step 2: REFLECT

After you answer AI’s questions, don’t let it run off to generate yet. Run this first:

Before you generate anything, tell me:

1. What assumptions are you making that I didn't explicitly confirm?
2. Based on my answers, what's your current understanding of what matters most?
3. What are you still uncertain about that you didn't ask?

This surfaces the hidden nonsense.

Real example: I asked AI to help with a newsletter about AI collaboration. (The irony isn’t lost on me. Nothing is lost on me anymore. I’ve stared into the recursive abyss and it wrote me a pretty decent outline.)

It asked solid questions. Good tone inquiries. Reasonable audience questions. I answered everything.

Then I ran the reflect prompt.

AI revealed: “I’m assuming your goal is to convince readers that AI is valuable and worth using.”

Wrong. My readers already use AI daily. They’re not skeptics who need converting (they’re past that existential crisis). They want better methods, not philosophical permission.

That assumption would have produced a completely different (and completely wrong) piece. A “why AI matters” manifesto when I needed a “how to stop hating AI” guide.

Thirty seconds of reflection. Saved me from publishing something that would make my own readers roll their eyes.

(I’ve published enough eye-roll-worthy content in my career. I’m trying to retire from that particular sport.)

Step 3: CORRECT

After AI generates, don’t just accept or reject. Do what you’d do with a junior team member who has potential but keeps making the same mistakes: give specific feedback that teaches the principle behind the correction.

Here's what I'm changing and why:

[Edit]: Changed "leverage AI capabilities" to "use AI"
[Why]: I don't speak corporate. Neither does my audience. We're all exhausted.

[Edit]: Cut the three-sentence introduction, started with the core idea
[Why]: I don't warm up. I start in the middle. Readers can catch up.

[Edit]: Added the example about the marketing email disaster
[Why]: Abstract advice is forgettable. Specific disasters are memorable.

What patterns do you notice in my corrections?

That last question forces AI to synthesize. “Don’t say leverage” becomes “this person hates business jargon.” One correction teaches a reusable principle.

Over time, this compounds. After several correction loops, AI starts anticipating your preferences. It asks better questions. Avoids your known triggers. Learns what actually helps versus what just sounds helpful.

That’s not prompting anymore.

That’s actual collaboration.

(Which is what we were promised, incidentally, before the “one-click content” crowd got their marketing budgets. But I digress.)

---

The Math (For Those Who Need It)

“This sounds like more work.”

It is. Up front.

Five minutes of intake versus zero minutes of intake.

But here’s what you skip:

• The “no, that’s not what I meant” loop (3 rounds minimum, each more soul-crushing than the last)

• Rewriting 70% of AI’s confident guess while questioning your career choices

• The frustration spiral where you start editing and end up deleting everything because it’s faster to start over

Five minutes of onboarding versus twenty minutes of cleanup.

And the onboarding compounds. Every correction you document teaches AI your patterns for next time. The new hire gets better. The Magic 8-Ball just keeps floating the same twenty generic answers forever.

Signs point to yes that you’re wasting your time with that approach.

---

The Quick Reference Card

Because you’re going to forget this in three days and need to look it up:

FEED:

[Your task]

Before generating anything, ask me 3-5 clarifying questions about:
- Who exactly will read this and what they already know
- What outcome I actually want (not what you assume I want)
- Constraints I haven't mentioned but will definitely be mad about later
- What "success" looks like vs. "this made everything worse"

Don't produce anything until I've answered. Your questions should help you understand my intent, not demonstrate your knowledge.

REFLECT:

Before you generate, tell me:
1. What assumptions are you making that I didn't explicitly confirm?
2. What's your current understanding of what matters most?
3. What are you still uncertain about?

CORRECT:

Here's what I'm changing and why:

[Edit]: [what you changed]
[Why]: [the principle behind it]

What patterns do you notice?

---

The Uncomfortable Truth

Here’s what most AI advice won’t tell you:

The problem isn’t AI. The problem is we’ve been trained to treat AI like a mystical oracle instead of a capable-but-clueless collaborator.

We’ve been told “just use this prompt” like there’s a magic formula that works for everyone. (There isn’t. That’s the point. That’s always been the point. If the same prompt worked for everyone, everyone’s content would sound the same, and... gestures vaguely at the internet... oh. Right.)

We’ve been sold automation when what we needed was collaboration.

The calibration loop isn’t sexy. It won’t promise 10X by Friday—but it will cut cleanup by half.

It works because it treats AI like what it actually is: a powerful tool that needs context to be useful. Not a fortune-telling orb. Not a replacement for thinking. Not a shortcut around the hard work of knowing what you actually want.

You wouldn’t skip the intake conversation with a brilliant new hire. You’d be insane to skip it.

Stop skipping it with AI.

---

The 8-Ball approach asks AI to guess. The onboarding approach teaches it to ask.

Same tools. Different relationship. Wildly different results.

Try this on your next AI task. Run the loop once. See what changes when AI stops floating random answers and starts having an actual conversation.

What’s the most spectacularly wrong output AI has ever confidently handed you? I collect these stories like some people collect Labubus. (And with similar regret.) Drop it in the comments.