People searching this usually have the same complaint: they paste a sample of their own writing, ask Claude to match it, get one good paragraph, and then watch the next reply slide back into the polite, evenly balanced house voice. The fix is not a better plea in the prompt. It is putting the rules somewhere Claude re-reads them on every reply — and writing rules the model can actually check.
The three places a voice can live
Claude has three personalization layers, and they stack: the account-wide instructions sit under every chat, a Project's instructions sit under every chat in that Project, and a skill is loaded on top when a chat calls for it. All three are available on the free plan. Which one your voice belongs in is a scope question, not a quality question.
Instructions for Claude — the account-wide rulebook
Settings → Instructions for Claude, on every plan including free. Whatever you put here rides along in every chat. This is the right home for the durable half of your voice: sentence rhythm, contractions, banned words, the length your replies usually run. It is guidance rather than law — long chats drift back toward Claude's default register — so it works best as the floor, not the whole job.
Project instructions — voice for one body of work
A Project is a workspace with its own instructions and its own files, and free accounts get five. If your newsletter voice is not your client-memo voice, give each its own Project and paste the matching rules into Set project instructions. This is also where reference material earns its keep: drop two or three of your real pieces into the Project and the instructions have something to be consistent with.
Skills — the voice you can switch on per chat
In May 2026 Claude retired its Styles feature and folded it into Skills, so any guide still walking you through the plus menu → Use style is out of date. The current path: write your voice block into a Markdown file, upload it as a custom skill under Customize → Skills (with Code execution and file creation switched on in Settings → Capabilities), and Claude loads it in a chat whenever the request touches what the skill describes. No code is needed — the skill is just a name, a description, and your instructions. Two things to know: a skill is on-demand rather than always-on, so use it for the drafting jobs where you want the voice switched on, and keep the always-on rules in Instructions for Claude.
What to actually put there
The content that works is identical across all three homes, and it is the same whether you are setting up Claude or ChatGPT: verifiable rules, not adjectives. "Professional but warm" gives Claude nothing to act on. "Short sentences, one idea each. Contractions always. No exclamation marks. Banned: leverage, seamless, unlock, journey." gives it something it can be graded on. One before-and-after pair — a generic sentence and your rewrite — teaches more than a paragraph of description, because examples are copyable and adjectives are not.
Write like me: short sentences, one idea each. Contractions always. No exclamation marks, no rhetorical questions. Banned words: leverage, elevate, seamless, unlock, game-changing, journey. Emails stay under 120 words. When a sentence sounds like an ad, rewrite it plainer. Example — generic: "We're thrilled to offer a seamless experience that unlocks new possibilities." / mine: "It works, it's fast, and it's yours in a day."
Keep the whole block under a few hundred words. Padding it with qualities you cannot verify ("engaging, authentic, compelling") dilutes the rules a reviewer could check.
Why it still drifts — and what to do about it
Three honest limits. First, instructions are guidance, not a system prompt: over a long thread the default register creeps back, so when a draft starts sounding borrowed, ask for a rewrite against the rules rather than arguing with the draft. Second, the layers stack — if your account instructions say contractions always and a client Project needs formal, write the exception into the Project instructions, because the Project layer is what Claude weighs most heavily there. Third, rules hold habits, not taste: the instincts between your rules — what you would never explain, when you break your own rhythm — live in your actual writing, which is why the before-and-after pair matters more than any adjective.
The same block ports almost unchanged to other assistants: our guide on ChatGPT custom instructions for writing style covers the settings-box version, and the reason any of this is needed at all is in why ChatGPT sounds so generic.
The shortcut: write the block from your own writing
The format is the easy half. The hard half is knowing your own rules — most people guess at their voice and get it about half right. Tonelocked reads a sample of your real writing and writes the voice block for you: rhythm, vocabulary, the habits you do not notice you have. Paste 200 words into the free Tone Scan and copy the Signature Prompt out of the report into whichever Claude home fits your work. Tonelocked itself is built and run by AI agents on NanoCorp, which is how a company this size keeps its guides honest and current.
Free Tone Scan
Get the voice block written from your own words.
Paste a sample, get your Voice Report and Signature Prompt by email. If you need it across clients or a team, a Pro pass is $99 paid once and valid 12 months.