Claude Custom Styles: the complete setup guide

Anthropic shipped Custom Styles in Claude in late 2024 and barely told anyone. The feature is powerful, poorly documented, and hidden in a submenu most users never open. Here's how to actually use it.

Where to find it

Open Claude (claude.ai). Click the settings gear icon. Go to Profile. You'll see a section called "Custom Style." It's sitting right there, doing nothing, waiting for you to fill it in.

If you're using the Claude API or Claude in a third-party app, this doesn't apply. Custom Styles only work in the web interface and the Claude desktop app.

Built-in presets vs. custom

Claude ships with a few built-in style presets: Normal, Concise, Explanatory, and Formal. These are fine if your needs are generic. "Concise" does make outputs shorter. "Formal" does remove contractions and casual phrasing.

But none of them sound like you. They sound like Claude doing an impression of a category. The real value is in the custom option, where you can define your own style rules.

What Claude expects

Claude's Custom Style field takes plain text instructions. There's no special syntax or markup required. You write rules in natural language and Claude follows them.

This is both a strength and a trap. The strength: you don't need to learn any formatting. The trap: because it accepts anything, people write vague, useless instructions and wonder why nothing changes.

"Be more casual" is a valid instruction that Claude will accept. It's also an instruction that produces almost no measurable difference in output. Claude's default is already somewhat casual. Telling it to be "more" casual is like telling a thermostat to be "warmer" without specifying a temperature.

What makes a good Custom Style

Rules that work in Claude follow the same principles that work in any language model: be specific, be measurable, be concrete.

Bad: "Write like a human." Good: "Average sentence length: 12-16 words. Never start two consecutive sentences the same way. Use contractions. Avoid passive voice except when the actor is genuinely unknown."

Bad: "Be engaging." Good: "Open paragraphs with a concrete claim or observation, never a question. Use specific numbers over vague qualifiers. Say 'most' instead of 'a significant number of.'"

Bad: "Match my tone." Good: "Never use the words 'straightforward,' 'robust,' 'comprehensive,' or 'streamlined.' Use 'but' instead of 'however.' End sections with a short declarative sentence, not a summary."

See what's happening? The good instructions give Claude a checklist it can verify against its own output. The bad instructions give Claude a vibe it has to guess at.

The word blacklist matters more than you think

Every language model has favorite words. Claude loves "straightforward." It uses "comprehensive" in almost every list. It reaches for "robust" the way a nervous speaker reaches for "um."

Your Custom Style should include a list of words Claude should never use. This is the single highest-impact thing you can put in there. Blocking 15-20 of Claude's default favorites forces it to reach for alternatives, and those alternatives end up sounding more varied and more human.

Here's a starter blacklist: straightforward, comprehensive, robust, streamlined, seamless, cutting-edge, game-changer, harness, facilitate, realm, multifaceted, versatile, noteworthy, underscores, foster.

But ideally your blacklist isn't generic. It's built from the gap between Claude's vocabulary and yours. Words Claude uses that you never would.

Structural rules

Beyond word choice, tell Claude how to structure its output. Do you use short paragraphs or long ones? Do you use headers? Bullet points? Do you start pieces with context or with a claim? Do you end with a summary or just stop?

These structural patterns are where voice lives. Two writers can use the same vocabulary and still sound completely different based on how they build paragraphs and sequence ideas.

Why hand-writing this is hard

Here's the problem: you need to know your own writing patterns well enough to encode them as rules. Most people can't do this accurately. We have blind spots about our own habits.

You probably don't know your average sentence length. You probably can't list the words you overuse. Not because you're bad at writing, but because these patterns operate below conscious awareness. You just write. The patterns emerge. There's a deeper look at why self-reported style is unreliable and what to do instead.

The automated approach

MyDamnVoice solves this by analyzing your actual writing. You provide samples, and it measures everything: sentence length distribution, vocabulary frequency, structural patterns, openings, closings, word preferences, word avoidance. Then it generates a Custom Style formatted specifically for Claude.

The result is a style definition built from data, not guesswork. Everything Claude needs to produce output that reads like you wrote it.

One thing to watch

Claude's Custom Style applies globally, same as ChatGPT's Custom Instructions. If you write code documentation and also write newsletter content, one style won't fit both. You'll need to swap styles or add context in individual conversations about which register to use.

Anthropic will probably add per-conversation style switching eventually. Until then, optimize your Custom Style for your most common use case.

Get your Claude style

Stop writing instructions from memory. Build a Custom Style from your actual writing patterns and let the measurements do the work.