Last week I asked ChatGPT a question I was sure was easy. I print a batch of desk gifts where a few positions on the plate get skipped, and I’d marked most of the skips by hand on a little grid. I missed a couple. So I handed it two pictures, my grid and a photo of the actual print bed, and asked which ones I’d missed.
It thought about it for four minutes. Then it handed my grid back with a block of about ten boxes painted red. Confident. Clean. Completely wrong. I counted the bed myself, top-left, left to right, row by row, and I’d missed exactly one.
That gap, ten confident red boxes versus one real answer, is worth understanding, because it shows up everywhere makers use AI. Once you see why it happens, you can stop it with a single change to how you ask.
Why ChatGPT gives confident wrong answers
When I asked it why it blew such a simple task, its own explanation was the lesson:
“I treated it like a visual pattern-matching problem instead of a counting/indexing problem… I inferred the answer from shape/area, and that is a bad method for this kind of task.”
In plain terms: it eyeballed the rough shape of the gap and guessed, instead of counting slot by slot. And here’s the part that matters for your shop. It does this with words too, not just images. Ask it to total an order, reconcile an inventory list, check which SKUs are missing, or count how many of something sit in a photo, and it will often pattern-match a plausible answer rather than do the boring discrete work. Then it delivers that guess in the same calm, certain voice it uses when it’s right.
That confidence is the trap. A wrong answer that sounds unsure gets double-checked. A wrong answer that sounds finished gets pasted straight into your listing, your invoice, or your production run. As I’ve written before, AI has gotten good enough that the lazy answer looks done, so you never find out what it missed.
So can you actually trust ChatGPT?
Yes, but conditionally, and the condition is the whole game. Trust depends on the kind of task:
- Fuzzy, generative work is where it shines and you can lean on it: drafting a product description, brainstorming names, rewriting a clunky paragraph, turning rough notes into a friendly email. There’s no single right answer, so “plausible and good” is exactly what you want.
- Exact, checkable work is where it quietly fails: counting, math, matching items to a list, dates and deadlines, anything with one correct answer. Here “plausible” is not good enough, and plausible is its default setting.
So the honest answer isn’t “trust it” or “don’t.” It’s: trust it for the fuzzy stuff, and for the exact stuff, make it prove its work before you believe it. Which brings us to the fix.
How to actually prompt ChatGPT for your maker business
The mistake most makers make is treating ChatGPT like a search box: type a question, take the answer. For exact tasks you have to treat it more like a sharp new assistant on their first day, someone you trust but still ask to show their work. Four habits do almost all of it:
- Tell it what kind of problem it is. “This is a counting task, not a guess-the-shape task.” Naming the method up front stops it from defaulting to the lazy one.
- Make it show its work before the answer. “Number every item, then compare, then tell me.” When it has to count out loud, it actually counts.
- Don’t let it generate before it verifies. AI image tools especially cannot count. Ask one to “mark the right boxes” and you get a pretty, made-up diagram. Get the answer confirmed first, then make the picture.
- Don’t reach for a bigger model when the fix is a better instruction. I asked whether a smarter setting would have gotten it right. Its honest answer was maybe, but probably not, because the problem was framing, not horsepower. This is the difference between pulling a slot-machine handle and hoping, and handing the AI an actual system you can repeat. The system is the part worth learning.
The one prompting fix, ready to copy
Next time you hand ChatGPT something with a single right answer, paste this before your question:
This is a counting problem, not a spot-the-difference problem. Number every item in a fixed order, left to right and top to bottom. Compare each position against what’s already marked. Show me your count. Then tell me only the specific positions I missed, and don’t generate any image until the count is verified.
Swap “positions” for whatever you’re checking. The shape of it works anywhere exactness matters:
- Inventory: “Number every SKU in this list and tell me which ones aren’t in the photo.”
- Orders: “Add these line items one at a time, show each running total, then give me the final.”
- Listings: “List every variant I gave you, then flag any I haven’t written a description for yet.”
That one paragraph would have turned my four-minute wrong answer into a ten-second right one. It forces the AI to do what your eyes already know how to do.
Count the slots, don’t guess the shape. Use the speed for the fuzzy work, and keep your maker brain in front of anything that has to be exact. That’s not being anti-AI. It’s just knowing which jobs it’s actually good at.
Want this kind of prompting move the week it lands, plus the workflows that make AI pull real weight in your shop? Come build with us in the Maker Growth Hub. We’d rather hand you the system than the slot machine. And if you’re weighing whether guidance like this is worth paying for, here’s an honest take on maker coaching.
Keep reading
- The AI stack for a maker business: what you actually need (and what’s hype)
- My AI Agents Sent a Newsletter to 655 People. Nobody Asked Me.
- The 3 AI workflows every maker should set up first
Ready to put this to work? Join the Maker Growth Hub.