Why AI Keeps Giving You Mediocre Answers
9 min read
Dev runs two pizza shops in Columbus, Ohio. Last March he opened ChatGPT for the first time, typed "write a social media post for my pizza restaurant," and waited to be impressed.
What came back: "Craving something delicious? Come on down to our pizzeria for hot, fresh pizza made with love! Perfect for family night or a quick bite with friends. See you soon!" There was an emoji. Possibly three.
Dev read it twice, closed the tab, and told his shift manager the AI thing was overhyped. Then he went back to writing posts himself at 11 p.m. after close — which was the exact chore he'd been trying to escape.
Here's what actually happened: the AI did precisely what he asked. That's the problem.
The average answer problem
Tools like Claude and ChatGPT are prediction machines. They've read a staggering slice of the internet, and when you ask for something, they generate the most statistically likely response to your exact words. Every detail you leave out — who it's for, what it's selling, how you talk — gets filled in with the most common choice found across millions of similar requests.
The most common audience for "a pizza restaurant post" is nobody in particular. The most common tone is chirpy marketing-speak. The most common offer is no offer at all. Stack up enough of those safe, middle-of-the-road guesses and you get the mathematical average of every pizza post ever written.
Call it the average answer problem. A vague prompt doesn't produce a bad answer. It produces the average one — and average, by definition, sounds like everyone else.
Once you see this, disappointing AI output reads differently. It stops being "the AI is dumb" and becomes "which details did I leave blank?" Usually, several.
What Dev actually needed
Dev didn't need a post about pizza in general. Tuesday nights at his Westgate location were pulling about 40 percent of Friday's volume, and he'd just launched a $12 large two-topping deal to fix that. His audience was local families who already follow the page. And his regulars expect one groan-worthy pizza pun per post — it's become his signature.
None of that lived inside "write a social media post for my pizza restaurant." So none of it showed up.
| Weak ask | Strong ask | |
|---|---|---|
| Audience | Anyone, which means no one | Families near the Westgate store who follow the page |
| Goal | Unstated | Fill slow Tuesday nights |
| Details | None | $12 large two-topping, Tuesdays only, dine-in or carryout |
| Voice | Default marketing chirp | Casual, ends with one bad pizza pun |
| Result | Could be any pizzeria on Earth | Could only be Dev's shop |
Look at the same request, before and after:
Weak:
Write a social media post for my pizza restaurant.
Strong:
Write a Facebook post for my pizza shop's Westgate location in
Columbus, Ohio. Audience: local families who already follow the
page. Goal: fill slow Tuesday nights. The offer: $12 large
two-topping pizza, Tuesdays only, dine-in or carryout. Keep it
under 80 words, casual, and end with one groan-worthy pizza pun.Notice there's nothing clever in the strong version. No secret keywords, no technical tricks. It's just decisions Dev had already made, written down. The model didn't get smarter between those two prompts. The ask got clearer.
A skill, not a talent
Some founders conclude they're "not AI people," the way some people decide they're not math people. Wrong lesson. Nobody is born knowing how to brief a machine — but you already know how to brief a person, and that covers most of the skill.
Think about the best delegators you've worked with. They aren't smarter than everyone else. They've simply stopped assuming other people can read their minds. Working with AI is the same discipline, minus the part where the other party gets tired of your detail.
The rest of this chapter builds that discipline into a habit, and in chapter two we'll turn it into a repeatable structure called The Five-Part Prompt — a skeleton you can fill in for almost any business task in under two minutes.
For now, one shift is enough: when the answer disappoints, don't blame the machine. Find the blank.
Try it now
Open Claude or ChatGPT and type the vaguest version of a real task — something like "write a promo post for my business." Read the answer and count how many details it guessed wrong. Then ask again, but first jot down four things: who it's for, what you want to happen, one concrete detail (a price, a date, a product name), and how you'd say it out loud to a customer. Compare the two answers side by side. Five minutes, total. You'll never send the vague version again.