How to Train Claude to Match Your Writing Voice (and Keep It Consistent)

You ask Claude to write something, and the output is competent, well-structured, and sounds like every other AI-generated paragraph on the internet. It is not wrong. It is just not you. You spend twenty minutes rewriting it and wonder whether the AI saved you any time at all. The problem is not Claude's writing ability — it is that Claude has no model of your specific voice to work from.

Why do voice descriptions fail?

The first instinct is to describe your voice in adjectives: "friendly but professional, concise, direct." These words tell Claude almost nothing actionable. "Professional" can mean anything from a legal brief to a marketing email. "Concise" might mean three sentences or three paragraphs depending on the context. Claude interprets these adjectives using its training data, which produces the average of all professional writing it has ever seen — not your voice specifically.

The second common approach is pasting a sample and saying "write like this." This works better initially but degrades quickly. Without structural rules to anchor it, Claude's attention to the sample weakens over the course of a long output or across multiple conversations. The voice drifts back toward generic AI prose. By the third paragraph of a long document, the sample's influence has faded significantly.

The approach that actually works combines three elements: structural rules that define what your voice does mechanically, writing samples that demonstrate what it sounds like in practice, and negative constraints that specify what it never does. All three are needed. Rules without examples are abstract. Examples without rules are fragile. And without constraints, Claude fills in gaps with its default patterns.

How do you extract structural rules from your own writing?

Analyse five to ten pieces of your best writing — documents, emails, reports, or articles you are genuinely proud of. Look for the structural patterns, not the content. What is your average sentence length? How long are your paragraphs? How do you open sections — with a statement, a question, or context-setting? How do you close them — with a conclusion, a call to action, or a transition? Do you use contractions? Do you use first person? How often do you use lists versus prose?

You can do this analysis manually or paste your samples into Claude and ask it to identify the patterns. Either way, encode the findings as explicit instructions. "Average twelve words per sentence. Paragraphs of two to four sentences. Open paragraphs with concrete statements, not questions. Use contractions in informal contexts. Never exceed four items in a bulleted list. Close sections with a forward-looking statement."

The more specific and structural your rules, the more consistently Claude can follow them. "Be concise" is unenforceable. "No sentence longer than twenty words" is a rule Claude can apply on every output.

How should you use writing samples to train Claude's voice?

Upload three to five pieces of writing that represent your voice at its best, ideally across different formats — a client email, a report section, a strategy document, an internal update. Place these in a Claude Project or reference them in a Skill file. The samples serve as a calibration reference that Claude uses alongside your structural rules.

The key is selecting samples that genuinely represent how you write at your best, not how you write when rushed or uninspired. If your samples contain inconsistencies — one formal, one casual, one with long sentences, one with short — Claude will average them and produce something that matches none of them. Curate deliberately. If you write differently for different audiences, create separate sample sets for each and use conditional rules to switch between them.

Why are negative constraints more powerful than positive rules?

This is the most valuable component of voice training and the one most people skip entirely. Build an explicit list of words, phrases, and patterns that your voice never uses. Never use "synergy," "leverage," or "streamline." Do not start paragraphs with "In today's fast-paced world" or "It's no secret that." Avoid rhetorical questions in professional writing. Never use "I hope this email finds you well." Do not use exclamation marks in client communication.

Negative constraints are more powerful than positive rules for two reasons. First, they are more specific — "never use this word" is binary and unambiguous, while "be direct" is open to interpretation. Second, they address Claude's most common failure mode, which is defaulting to generic AI-sounding phrases when it lacks specific guidance. A constraint list acts as a filter that catches these defaults before they reach the output.

A good constraint list has twenty to fifty entries. It grows over time as you notice new patterns in Claude's output that do not match your voice. Every time you edit something Claude produces, ask yourself: is there a rule that would have prevented this? If so, add it.

How do you maintain voice consistency across conversations?

Build voice as a Skill, not a one-off instruction. Skills persist across conversations and load automatically when they match the task. A voice Skill that contains your rules, references your samples, and encodes your constraints will produce consistent output whether you are drafting an email on Monday or writing a report on Friday. The voice does not drift because the Skill does not change.

For multi-audience work, build conditional rules within the Skill. "When writing for clients, use formal register and avoid contractions. When writing for internal colleagues, use conversational register and contractions are fine. When writing for leadership, lead with the conclusion and keep paragraphs to two sentences." These conditionals allow a single voice Skill to adapt without losing your core patterns.

MyOS builds this entire system through guided extraction — the Voice module analyses how you actually write and generates a Skill file with rules, examples, and constraints calibrated to your specific patterns.

For the broader personalisation system that voice plugs into, see our guide on building a personalised AI system. If you are experiencing drift and inconsistency, our troubleshooting guide covers the fixes. Voice training is just one part of a full Claude setup — once your voice is locked in, layer in workflow Skills for your recurring tasks.