I’ve lost count of how many prompts I’ve seen that read like a legal disclaimer. Thirty lines of “don’t do this, avoid that,” and the draft still comes back sounding like every other AI paragraph on the internet. So when I spotted this post from a LinkedIn creator who fixed the problem with a single file instead of a pile of restrictions, I had to pass it along.
The author watched the pattern play out on their own team constantly. Someone would spend ten minutes loading a prompt with rules. Don’t use buzzwords. Don’t open with a tired cliché line. Don’t sound like AI. Then the output would land, and it still sounded like AI. Every single time.
🧠 Why rules-only prompts keep failing
The original poster boils the reason down to one line, and honestly it rewired how I think about prompting: rules tell the model what to avoid. They don’t show it what to do.
Think about it from the model’s side. You’ve handed it a list of thirty exits and zero destinations. It knows the doors it can’t walk through, but it has no idea where you actually want it to go. So it defaults to the safest, most average version of writing it knows. That’s the generic voice you keep getting. No examples, no voice, no reference point means no target to aim at.
That one distinction, avoid versus do, is what changed how the author’s whole team prompts now.
📁 The fix: a reference file instead of a rulebook
The swap this savvy professional made is simple. Instead of restrictions, you give the model a reference file. Ten to fifteen pieces of your actual writing. Real posts, real emails, real threads. The ones you’re genuinely proud of.
The model reads them and picks up your sentence rhythm, your vocabulary, how you open, how you close. Then you add just one line to the prompt:
“Apply everything you learn as a writing rulebook.”
That’s it. No thirty-line list of bans. The examples do the heavy lifting, because the model now has something concrete to match instead of a fog of things to dodge.
🛠️ How to build a reference file that works
Here’s the exact build the expert shared, and I’ve added a quick note on why each step matters so you can adapt it to your own situation.
- Save 10 to 15 pieces of writing you’re proud of. Posts, emails, threads. The rationale: the model learns from patterns, and a handful of samples isn’t enough to spot them. Fifteen gives it room to see what repeats in your style versus what was a one-off.
- Paste them all into one document. One file, not fifteen. This keeps the upload to a single action and makes it easy to reuse across tools and sessions.
- Label each piece with a one-line note on why it worked. Something like “High engagement, conversational tone” or “Short sentences, strong opening.” This tells the model which traits to prioritize, so it’s not just copying your words, it’s copying the moves behind them.
- Add a short section at the top describing your voice in plain language. Five or six sentences max. This is the summary layer. It gives the model a mental frame before it reads the examples, which makes the patterns click faster.
- Upload this file every time you start a new writing session with AI. Most tools don’t remember past sessions reliably, so make the upload part of your routine. Thirty seconds, every time.
- Tell the model to study it before writing anything. Don’t let it skim and start drafting. Ask it to read the file first, then ask you clarifying questions before it produces a single word. That pause is where the quality jumps.
⚡ The before and after
The contributor describes the shift as dramatic, and I believe it.
- 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.
That’s not just a time saving. It’s a completely different relationship with the tool. You stop policing it and start teaching it.
💡 A few ways I’d push this further
The core method is solid on its own, but here’s how I’d apply it in practice:
- Build one file per format. Your newsletter voice and your LinkedIn voice probably aren’t identical. A separate reference file for each keeps the model from blending them.
- Include a couple of “almost but not quite” examples. Label them honestly: “Too formal, this is what I sound like on a bad day.” Contrast helps the model sharpen its sense of your good days.
- Refresh the file every quarter. Your writing evolves. Swap out older pieces for recent ones so the model tracks where you’re headed, not where you were.
- Keep the clarifying-questions step. It’s easy to skip when you’re in a hurry, but it’s the cheapest quality control you’ll ever get.
This is one of those ideas that seems obvious the second you hear it, and yet almost nobody does it. Rules feel productive. Examples actually work.
The author closes with a great question worth asking yourself: are you already using a reference file, or still working from rules? If you’ve got a teammate still running rules-only prompts, send this their way. And check out the full LinkedIn post for the complete breakdown and the discussion in the comments.