You can smell an AI-written product review from the first sentence. "In today's fast-paced world, finding the right product can be a game-changer." Nobody talks like that. Nobody who has actually used a slow-feeder bowl for three months opens with a thesis statement.
I run a pet-gear review site, and I wanted to use AI to speed up drafting without producing that. What I found is that the prompt is the entire job. I ended up rewriting one review-drafting prompt around fifteen times, and the interesting part wasn't getting the AI to write more — it was getting it to write less like an AI. Here's what actually moved the needle, and the hard ceiling I kept hitting.
Why the first prompt failed
The naive version is the one everyone starts with:
Write a review of [product]. Make it sound natural and helpful.
What comes back is fluent and completely hollow. It has the shape of a review with none of the substance: sweeping adjectives, zero specifics, a conclusion that could apply to any product in the category. "Make it sound natural" does nothing, because the model's default register already is what it thinks natural sounds like — and its idea of natural is a marketing brochure.
The problem isn't fluency. The model has plenty. The problem is that generic prose is its lowest-energy output, and a vague prompt lets it sit there.
What actually worked
The fixes were all about removing the model's room to be generic.
A banned-phrase list. I keep an explicit blocklist in the prompt — "game-changer," "in today's world," "look no further," "when it comes to," "elevate your," "whether you're a... or a..." Naming the clichés is far more effective than asking for "no clichés," because the model can't reliably detect its own tells; it can follow a literal list.
Forced structure with required slots. Instead of "write a review," the prompt demands specific components: the use case it's actually for, who it's not for, and at least one concrete drawback. The single biggest jump in believability came from forcing a real negative. Reviews with no downside read as ads; humans instinctively distrust them. A required "here's what annoyed me" slot breaks the brochure tone instantly.
Specificity over judgment. I forbid bare evaluations. The model can't write "it works great" — it has to write under what conditions it behaves how. "Great for a strong puller" is a claim; "it works great" is filler. Forcing the conditional drags the text toward something testable.
Constrained voice and density. First person, restrained, no encyclopedia tone, no summarizing the entire product category before getting to the point. I cap how much throat-clearing it's allowed before the first concrete observation.
Each of these is a small rule. Stacked together, they pull the output out of brochure-land and into something that at least reads like a person wrote it.
(For the mechanical drafting I leaned on Claude — it's good at following a strict, rule-heavy prompt like this. The prompt design was the work; the generation was the easy part.)
The ceiling no prompt gets past
Here's where I have to be honest, because it's the whole point.
Every technique above improves how the review reads. None of them improves whether it's true. You can make AI text sound like a person who used the product for three months. You cannot make it into a person who used the product for three months.
The things that make a review actually worth reading — how it held up after real use, the specific way it failed that the product page never mentions, the exact measurement, the photo of the thing dirty and dented on your kitchen floor — the model doesn't have any of that. If you let it fill those gaps, it invents them, and now you're not writing a thin review, you're writing a confidently false one. That's worse.
So the line I settled on: AI drafts and structures; a human supplies the experience. The prompt produces a clean, specific, human-sounding skeleton. Then someone who has actually handled the product fills in the parts only use can produce — and cuts anything the draft asserted that isn't true. On a pet site especially, where people make decisions that affect an animal's health, getting that division wrong isn't a style problem.
This also happens to line up with where search is going. Google's "helpful content" push is explicitly hunting for first-hand experience and demoting content that's generated to rank. A pile of fluent, experience-free reviews is exactly the target. The fluency was never the moat. The experience is.
That's the workflow behind pawpry.com — AI for the draft, a real person for the parts that matter. The prompt engineering got me a better starting line, not a finish line.
The takeaway
If you're using AI to draft content, spend your effort on the prompt, not on editing the output into shape afterward — a good prompt prevents the AI smell, late editing only masks it. But know exactly what the prompt can and can't buy you. It buys readability. It does not buy experience, and the moment you ask it to fake experience, you've built the thing everyone — readers and search engines alike — is learning to filter out.
Solo indie dev, writing these up as I go. If you've found prompt tricks that kill the AI tone, I'll take them — and I'm curious where everyone else draws the human-in-the-loop line.
Top comments (0)