6 things AI power users never ask Claude to do, and what they ask instead
How to Stop Claude AI Sounding Like AI
There’s a tell. Most people who use Claude regularly have seen it without naming it: you ask for something, the output arrives fast, it’s technically correct, and something about it feels... assembled. Like it was built from parts rather than written by someone who cared.
The problem usually isn’t the prompt. It’s what you asked for.
Power users have worked out, mostly through throwing away a lot of drafts, that certain requests almost guarantee that assembled feeling. The ask itself is the problem. You’re pointing Claude at the part of its training where everything sounds the same.
Six of those requests. All common. All worth stopping.
1. “Write me an introduction”
The introduction is the part Claude is worst at, and it’s the first thing most people ask for.
Ask Claude to write an introduction and you’ll almost always get throat-clearing: a sentence that restates the topic, a sentence that explains why it matters, and a transition into the main content. It reads like a five-paragraph essay format got applied to whatever you’re writing.
What power users do instead: they write the introduction themselves, or they ask Claude to write the second paragraph first.
The reason this works is that Claude’s defaults for openings are pulled from an enormous body of content where openings are formulaic. The middle of a piece has more variation. Start there, and the opening you write to connect to it will feel more natural because you’ve already got something real to connect to.
If you want Claude to write an opener, give it the whole piece first and ask: “Write a one or two sentence opening that drops straight into the argument. No context-setting, no explaining what the piece is about. Assume the reader already knows why they’re here.”
That constraint breaks the formula.
That fix alone changes what you get back. The constraint at the end of that prompt (”no context-setting, no explaining what the piece is about”) is doing more work than most full rewrites would.
The next five mistakes are less obvious, and a couple of them are genuinely counterintuitive. Subscribers get:
Mistake 2 — “Make it more engaging”: Why vague quality instructions produce vague quality improvements, and the specific questions that actually fix flat writing.
Mistake 3 — “Give me 10 ideas”: The padding problem that gets worse the higher the number, and the single-question move that produces better output than any list.
Mistake 4 — “Write me a summary”: Why summary requests produce useless output, and how naming the use case and audience changes everything you get back.
Mistake 5 — “Does this sound okay?”: The feedback request that almost always produces reassurance with a garnish, and how to ask for actual criticism instead.
Mistake 6 — “Write it in my voice”: Why naming your voice gives Claude nothing to work from, and the showing-not-naming approach that produces output that sounds like a specific person wrote it.
Plus: the constraint swipe file, six phrases that replace vague instructions with something Claude can act on.