How you talk to Claude shapes everything you get back. The mechanism, the limits, and why this matters more if you care about voice.
You can tell what Claude was prompted with by reading the output. The terse three-sentence summary came from a terse three-sentence prompt. The bullet list came from "give me bullets." The 4,000-word generic essay came from "write a comprehensive guide." The horoscope-shaped brand voice came from "describe the voice."
Claude is a pattern matcher. Your input is part of the pattern. The output is the most likely continuation of what you wrote.
If you're a dry shit linear thinker who treats every Claude prompt like a function call, you're going to hate the output. Not because Claude is bad. Because Claude is doing exactly what you asked. Function-call inputs produce function-call outputs. The model can do prose. You didn't write a prompt that asked for prose.
Most "Claude isn't producing good output for me" complaints trace back to prompts that are too curt to give Claude anything to work with. The fix isn't a better model. It's a better input. The model is mirroring exactly what you gave it.
Search-query prompts produce search-query outputs. Conversational prompts produce conversational outputs. The lever is upstream of the model.
Claude is auto-completing your conversation. The most likely next message after "summarize this 5 bullets" is five terse bullets. The most likely next message after "I'm prepping a Q3 review for our exec team and I need to think through how to present a metric that went sideways, can you help me draft a way to talk about it that's honest without being defensive" is a draft that's honest without being defensive. Different shape, different output.
This isn't a personality trait of Claude. It's how the underlying mechanism works. The model picks up on register from your input and continues it. Same way it picks up on tone, formality, slang, length of sentences, density of jargon. Anything in your prompt that's a pattern gets continued.
People who say "AI output sounds like AI" are usually correct, but they're wrong about the cause. The output sounds like AI because the prompt sounded like a search query. Search queries don't have tone. The output won't either.
"Give me 10 LinkedIn post ideas" produces ten LinkedIn-post-ideas-shaped outputs. Generic, interchangeable, identifiable.
"I'm trying to drive thinking about [specific topic] for [specific audience], and I want to test [specific angle]. Can we explore three drafts that approach it differently, and one you think I'd hate but is the strongest argument the other side has?" produces three drafts you can actually use, plus a fourth that sharpens your thinking.
The second prompt is longer. It's also one a colleague would write to another colleague. That's not a coincidence. That's the lever.
Not a god. Not a search engine. A capable person who needs context.
The amount of context you'd give a smart junior is roughly the amount Claude needs to do the job well. If you wouldn't ask a junior "give me a marketing strategy," don't ask Claude that. If you would ask a junior "look at these three competitors, here's what I think the white space might be, can you stress-test that hypothesis and surface one I missed," ask Claude that.
The junior framing also fixes the trust calibration. You don't ship the junior's first draft to the client. You read it, edit it, send back questions, ask for a second pass. Same pattern with Claude. The output is a draft you sharpen, not a deliverable you forward.
This is the limit. Being warm and contextual doesn't override structural problems with the prompt. A friendly request for "tell me about our brand voice" still gets adjective slop because the question doesn't have a refuse condition or an output spec. The fix for that isn't politeness. It's structure.
But the structural prompts work better when they're written conversationally too. "Pull six examples from these samples where the brand uses the voice well, then rewrite each one in the most common alternative phrasing. The contrast IS the guide" is a structural prompt written in human voice. It works better than the same prompt written as a JSON instruction set with bullet points and ALL CAPS section headers.
Structure tells Claude what the output should be. Register tells Claude how to write it. You need both. You can't substitute one for the other.
If your brand voice matters, this is the biggest lever you're not using.
A copywriter who writes prompts like Stack Overflow queries gets back content that sounds like Stack Overflow answers. Then they edit it for hours to sound human. That's the slop tax in real time.
A copywriter who writes prompts the way they actually think gets back content that sounds the way they think. Then they edit it for accuracy and structure. The voice is already there. The edit is shorter.
This post you're reading right now was drafted from a prompt that gave Claude specific phrases I'd actually use rather than generic alternatives. That single instruction shaped the draft. Not because Claude is being warm to me. Because Claude is doing what auto-complete does, and the input had specific language to autocomplete from.
Wild goose chase if you want one: scan the other posts on this site and try to guess which lines came from a prompt and which were typed by me. The honest answer is most of them are both. The posts that sound most like me had the sharpest constraints in the input. That's the thesis of this post applied to itself.
Try it now. Pick a marketing task you actually need to do this week. Write the prompt twice.
Version 1: terse, command-shaped. "Write me an email about X for Y audience."
Version 2: conversational, with context. "I'm trying to send X to Y and I'm worried about Z. Can we draft three approaches and stress-test which one would actually land with someone who's seen this kind of email a hundred times before?"
Read both outputs. The second one will be measurably better. Not because Claude is nicer to people who are nice. Because the second prompt gave Claude something to work with.
People are the whole point of this exercise. Claude is the tool. If you treat the tool like a search engine you'll get search engine output. If you treat it like a colleague you'll get colleague output. The choice is upstream of the model. It's in your fingers.
Each one is structured (refuse conditions, output specs, real source material) but written like a colleague brief, not a function call. That's why the outputs feel different from the listicle versions. Read them, run them, see if you notice the shift.
See the 4 free prompts →