I spent an embarrassing amount of time last year writing prompt rules. Don’t use buzzwords. Don’t open with a big sweeping statement. Don’t sound like a robot. My prompts slowly turned into legal documents. And the output still sounded like every other AI post floating around the internet.
Then I came across a post from an AI professional that explained exactly why that keeps happening. The fix took about thirty seconds to set up. I was genuinely annoyed at how simple it was.
Why rules-only prompts keep failing
The original poster puts it in one line: rules tell the model what to avoid, they don’t show it what to do.
Think about what you’re actually handing the model when you stack up 30 restrictions. A list of things to avoid. No examples. No voice. No reference point. You’ve drawn the fence, but you never showed it the house. So the model does the only thing it can: it writes something technically inside the fence and completely forgettable.
The author watched this play out on their own team constantly. Ten minutes loading a prompt with restrictions. Don’t use buzzwords. Don’t sound like AI. And the output would still sound like AI. Every single time.
Restrictions define the boundary. Examples define the target. You need the target.
The swap that changes the whole thing
Instead of rules, you give the model a reference file. Ten to fifteen pieces of your actual writing. Real posts, real emails, real threads. Ones you’re actually proud of.
The model reads them and picks up your sentence rhythm, your vocabulary, how you open, how you close. All the stuff you could never write into a rule because you don’t consciously know you’re doing it. Nobody thinks “I tend to open with a three-word sentence.” But your writing knows.
Then you write one line in the prompt:
“Apply everything you learn as a writing rulebook.”
That’s it. That’s the entire prompt change.
The exact build, step by step
Here’s the build this LinkedIn creator shared, with the reasoning behind why each step earns its spot:
- Save 10 to 15 pieces of writing you’re proud of. Posts, emails, threads. The range matters. Fewer than ten and the model latches onto the quirks of one piece instead of the pattern across all of them.
- Paste them all into one document. One file, one upload, one context. You’re not hunting through folders in the middle of a writing session, and the model sees everything at once instead of in fragments.
- Label each one with a one-line note on why it worked. “High engagement, conversational tone.” “Short sentences, strong opening.” This is the step most people skip, and it’s the one doing the heavy lifting. You’re telling the model what to notice, not just what to read.
- Add a short voice section at the top, five or six sentences max. Plain language. No jargon. This gives the model a frame before it starts reading, so the examples land as evidence for something instead of random samples.
- Upload the file every time you start a new writing session. Chats don’t carry memory the way you assume they do. Fresh session, fresh upload. It’s a ten second tax that saves you a bad first draft.
- Tell the model to study it before writing anything, and ask it to ask you clarifying questions first. This forces actual processing instead of a skim. The questions are a bonus signal: if it asks something smart, it read the file properly.
The before and after
The expert lays out the contrast cleanly, and it’s the part that stuck with me.
- Before: ten minutes writing restrictions into every prompt, still getting generic output.
- After: thirty seconds uploading one file, and the first draft sounds like you.
The time math alone is worth it. But the real shift is that you stop editing AI output back into your voice, which is where most of the hidden hours go. Rewriting a generic draft into something that sounds like you takes longer than writing it yourself. A reference file skips that whole round trip.
How I’d put this to work today
A few practical angles the post got me thinking about:
- Build one per format. Your LinkedIn voice and your client email voice aren’t the same. Two files, two rulebooks.
- Refresh it quarterly. Your writing improves. Swap in the newer pieces you’re proud of and drop the ones that feel dated.
- Share it with your team. If you’ve got a house voice, one reference file beats a 30 page brand guideline nobody reads. The examples do the explaining.
- Use it for editing too. Upload the file, paste a draft, and ask the model to flag anything that doesn’t match the reference. It catches things you’ve gone blind to.
- Include the misses on purpose. Add one or two pieces labeled “this one flopped, tone was too stiff.” Contrast teaches faster than a clean set of winners.
Why this fits where AI is heading
Models keep getting better at following instructions, and that’s exactly why rule-stacking feels productive while producing nothing. The constraint was never the model’s obedience. It was that you never gave it a target to aim at. Show, don’t tell has been writing advice for a century. Turns out it applies to prompting too.
I think this contributor nailed something most prompt advice misses. The best prompt isn’t a longer prompt. It’s a shorter prompt with better context attached.
The original post has the full breakdown plus an infographic version of the build. Worth a look if you want the whole thing in one visual. Go check it out, then go dig up ten pieces of your own writing.