How to Make Claude Write in Your Voice

A freelance ghostwriter working with a fintech CEO tested twelve different prompt formats before finding one that actually captured his voice instead of a smoothed-out, generic approximation of it. That's a lot of trial and error for something that seems like it should be simple. But most people learning how to make Claude write in your voice hit the same wall she did at first: they describe the voice with adjectives, and adjectives don't transfer style the way real examples do.

Before: The Weak Prompt

Here's what she started with:

Write in a confident, direct, slightly informal voice. The tone should 
be smart but approachable, not too corporate. Here's a topic: why most 
startups underprice their first product.

What this does: it hands Claude a set of style adjectives and asks it to interpret them, which means the output reflects Claude's default idea of "confident and direct" rather than this specific person's actual voice. Two different writers could both fit "confident, direct, slightly informal" and sound nothing alike.

Why It Fails

Adjectives describe a voice the way a police sketch describes a face — recognizable in broad strokes, useless for identifying the actual person. "Direct" could mean short sentences and blunt claims, or it could mean skipping pleasantries but writing in long, layered paragraphs. Claude has to guess which, and it guesses toward the most common pattern it's seen for that adjective, which is rarely the specific pattern this one person actually uses.

⚠️ Common mistake: describing a voice instead of showing it. If you can write three adjectives to describe someone's writing style, you can probably find three paragraphs of their actual writing instead — and the second one works dramatically better.

After: The Improved Prompt

Here are three paragraphs the client has previously published (voice 
samples below). Study the sentence structure, word choice, and how 
they transition between ideas — don't summarize the content, focus on 
the mechanics of how they write.

[Paragraph 1]
[Paragraph 2]
[Paragraph 3]

Now write a new piece on "why most startups underprice their first 
product" matching this voice. Specifically:
- Match sentence length variation (they mix short punchy sentences 
  with one longer one per paragraph)
- They rarely use rhetorical questions — avoid those
- They use one concrete number or example per main point, not 
  abstractions
- They end sections with a short, standalone sentence for emphasis

What this does: instead of asking Claude to interpret vague style words, it gives real samples and calls out the specific mechanical patterns to replicate, which is closer to what a human ghostwriter actually studies when learning to write as someone else.

⚡ Pro tip: pull the mechanical patterns out yourself first — sentence length, punctuation habits, how they open and close paragraphs — rather than asking Claude to "analyze the voice" as a separate step. Naming the patterns explicitly gets more consistent results than asking Claude to infer and apply them in one pass.

Breaking Down Each Element

The request to "focus on mechanics, not content" matters more than it looks. Without it, Claude sometimes pattern-matches on subject matter and tone in a way that drifts toward summarizing what the samples said rather than how they said it. Naming specific structural habits — sentence variation, rhetorical questions, closing style — turns a fuzzy imitation task into a checklist Claude can actually verify against as it writes.

A content marketer at a B2B insurance company used this exact structure to keep three different guest contributors sounding like themselves across ghostwritten LinkedIn posts, instead of all converging toward one "LinkedIn voice." A journalist turned newsletter writer used it to maintain her own voice when scaling from writing every issue herself to reviewing Claude-assisted first drafts — she found that without explicit mechanical notes, drafts drifted toward more balanced, hedged sentences than she'd ever actually write herself.

⚡ Pro tip: if you don't have three paragraphs of existing writing to pull from, record yourself talking about the topic for two minutes and transcribe it. Spoken patterns — where someone naturally pauses, what they emphasize, how they qualify a claim — translate into voice cues almost as well as written samples do.

Variations for Different Contexts

For a brand voice rather than an individual's, the process shifts slightly: instead of three paragraphs from one person, pull five or six sentences that a brand style guide already flags as "on voice," since brand voice tends to be more about consistent word choices (does the brand say "customers" or "clients"? "we" or the company name?) than about individual sentence rhythm.

For academic or technical writers moving into more casual content, it helps to explicitly name what to drop, not just what to add — "remove hedging phrases like 'it could be argued that,' keep technical precision but cut academic throat-clearing" gets a cleaner result than just asking for "more casual" writing, which tends to just add contractions on top of the same underlying sentence structure.

One pattern worth flagging: voice-matching degrades on longer pieces unless you re-anchor periodically. A newsletter writer found that by the third or fourth section of a long piece, Claude's output would drift back toward more generic phrasing, even with strong voice samples up front. Pasting one of the original voice samples again partway through a long draft — as a reminder, not a new instruction — reliably pulled the voice back into alignment.

One more variable worth testing deliberately: how many samples is enough. Two paragraphs sometimes isn't enough signal for Claude to distinguish real habits from one-off word choices; five or six paragraphs occasionally introduces enough inconsistency (since most people's writing varies somewhat by mood or context) that Claude ends up averaging across contradictory signals. Three to four paragraphs, ideally from similar contexts (all blog posts, or all emails, not a mix), tends to land in the sweet spot most consistently.

Save and Reuse This

The core structure — real samples, named mechanical patterns, explicit instruction to focus on mechanics over content — works for any voice-matching task, whether that's your own writing, a client's, or a brand's. What changes is just which samples you feed in and which patterns you name.

Since good voice samples take real effort to identify and describe well, it's worth keeping the full prompt — samples included — saved somewhere reusable rather than rebuilding it from scratch each time. PromptABCD works well for storing these voice profiles, so switching between three different clients' voices doesn't mean re-analyzing their writing from zero every single time.