I spent last year memorizing prompt “rules” like they were gospel. Add “step-by-step” here, stack a few examples there, tack on “right?” at the end for good measure. So when I scrolled past a LinkedIn post claiming that most of those habits are now making AI answers worse, I stopped cold.
The breakdown comes from an AI professional who tested the old-school prompting playbook against a stripped-down, natural approach. The results flipped a lot of what I thought I knew. I was genuinely surprised, because I’ve used some of these “bad” moves more often than I’d like to admit.
So let’s put the two styles head to head. Old way versus new way, tip by tip, with a clear pick at the end.
The core clash: over-engineered vs. natural
The old approach treats the prompt like a machine you have to program. Add scaffolding, add framing, add fake politeness. The new approach, according to the original poster, treats the prompt like a real question you’d ask a sharp colleague.
Here’s the pattern that jumped out at me first.
The “right?” trap. Ending a prompt with a leading tag like “…right?” quietly biases the model. The expert points out that AI gets swayed by how you ask more than by what it actually “knows.” Frame it as yes/no or nudge it toward an answer, and it will happily agree with you.
- Old way: “This approach is the best one, right?”
- New way: Ask the question naturally and avoid yes/no framing entirely.
Round two: “step-by-step” vs. just add your role
This is the one that stung. The creator admits they leaned on “step-by-step” constantly, and so did I. Turns out that stacking instructions like “gather information, devise a plan, answer step by step” can clutter the request instead of sharpening it.
The cleaner move is to drop the ritual and just tell the AI who it should be answering as.
Bad prompt: “[Your question] Gather information, devise a plan, answer step by step.”
Good prompt: “[Your question] . . . as a reliability engineer.”
Same question, a fraction of the words, and a role that actually steers the tone and depth. I love how simple that swap is.
Round three: confident answers vs. verified answers
Tip three from the post is a reality check. AI still hallucinates citations, and those sources sometimes flat-out do not exist. The original poster makes a blunt point: the only way to catch a fake link is to open the actual page, not the tidy summary the AI hands you.
And nobody is opening 210 sources to fact-check a single answer. That’s the honest tension here.
- Old habit: Trust the confident tone and the neat reference list.
- Better habit: Treat every citation as unverified until you click through to the real link.
Round four: perfect examples vs. a clear goal
Here’s where a lot of us pour our energy in the wrong bucket. The instinct is to hunt for the perfect examples to feed the model. The expert argues those examples can actually sabotage the output by boxing the AI into your sample cases instead of your real problem.
The fix is to spend that time specifying the goal clearly and handing over real context. Compare these two:
Bad prompt:
“You are a world-class newsletter strategist.
Example 1: [a solved case, written out]
Example 2: [a solved case, written out]
Find a plan, and answer step by step: How can I improve my newsletter?”
Good prompt:
“My newsletter open rate fell from [X]% to [Y]% over three months. I send one issue every Tuesday. I did not change the format, the send time, or the subject. What are the most likely causes?
Ask me for the data you need before you answer.”
See the difference? The good version trades fake expertise cosplay for real numbers, real constraints, and one brilliant closing move: telling the AI to ask you for missing data before it answers. That single line kills half the guesswork.
Quick side-by-side scorecard
- Leading tags (“right?”): Old = biased answers. New = neutral, natural questions win.
- “Step-by-step”: Old = cluttered instructions. New = question plus a short role wins.
- Citations: Old = trust the summary. New = open the real links, or don’t rely on them.
- Examples: Old = stuffing sample cases. New = clear goal plus real context wins.
My take and the recommendation
If I’m scoring this matchup, the natural approach wins on every round. The old playbook felt productive because it was more work, and more work feels like more control. But the expert’s point landed for me: extra scaffolding often just adds noise and bias.
So here’s the practical path I’m taking from this contributor’s findings. Ask like a human. Give real context and constraints instead of canned examples. Name the role in a few words rather than reciting “step by step.” And never trust a citation you haven’t opened yourself.
I think this is a real shift in how prompting works, and it’s refreshing because it asks for less effort, not more. The person who shared it packed even more tips and a full prompt template into the original.
Go check out the full LinkedIn post for the rest of the tips and the template. It’s worth the read.