Voice Fingerprint
marcusnelson/write-like-me-claude-skill
Ensine Claude a escrever como você, não como um profissional genérico.
Ver no GitHubSkill que ensina Claude a escrever com sua voz pessoal. Você envia amostras de texto, a ferramenta analisa padrões de escrita (ritmo, vocabulário, palavras evitadas) e cria um 'spec' estruturado que Claude usa para replicar seu estilo em todas as futuras tarefas.
Instalação
npx convez-ai marcusnelson/write-like-me-claude-skill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Gerar emails e propostas comerciais com voz autêntica
- Padronizar escrita de conteúdo mantendo estilo pessoal
- Refinar textos de marketing/redes sociais sem perder identidade
Veredito

Skill bem pensada que resolve um problema real: IA não consegue capturar voz autêntica sem referência explícita. O sistema de feedback contínuo e spec armazenado em memória é sólido, mas depende de amostras iniciais de qualidade e disposição do usuário em responder entrevistas estruturadas.
README
A tradução para o português deste README ainda está na fila.
Voice Fingerprint
Teach Claude how you write. Not how "a professional" writes. How you write.
AI writing tools are fluent. They're also generic. Ask five of them to write a pitch email and you'll get five versions of the same email — competent, forgettable, and nothing like you.
Voice Fingerprint fixes that. It reads your actual writing, identifies what makes it yours, interviews you about the choices you make on purpose, and produces a Voice Spec — a structured document that tells Claude exactly how to sound like you. Not a vague "tone: professional yet approachable." Concrete rules. Specific examples. A banned word list with reasons. The sentence rhythms you reach for. The clichés you've been quietly avoiding your entire career.
Then it gets better over time. Every correction ("I'd never say it that way") and every approval ("nailed it") feeds back into the spec. Your AI writing gets sharper the more you use it.
What It Actually Does
Reads your writing. Upload 3-8 samples — blog posts, emails, pitch decks, social posts, memos, whatever. Diversity matters more than volume. A pitch email and a LinkedIn post reveal more about your range than eight blog posts.
Finds the patterns you don't notice. Sentence length distribution. Whether you lean Anglo-Saxon or Latinate. How you open a piece. How you close one. Whether you use em-dashes for drama or parenthetical asides for texture.
Finds what's missing. This is the part most style tools skip. Voice Fingerprint cross-references your vocabulary against common cliché pools — consulting jargon, marketing buzzwords, tech filler — and flags what's conspicuously absent. If you've written 10,000 words of marketing copy and never once typed "synergy," "leverage," or "game-changer," that's not an accident. That's your voice.
Interviews you. Samples can't reveal everything. The skill asks targeted questions: what words you deliberately avoid and why (the why catches adjacent bans you haven't thought of yet), British or American English conventions, where you want controlled randomness, how your voice shifts by audience, and what kind of writing makes you cringe.
Builds a Voice Spec. A structured document with your five non-negotiable signature traits, mechanical rules, audience modes, a tiered banned list, anti-patterns, and example rewrites showing the transformation from generic to you. Stored in Claude's memory so every future conversation inherits it.
Learns from corrections. Say "that doesn't sound like me" and the skill identifies what went wrong, generalizes the pattern, and proposes a spec update. Say "that's perfect" and the skill strengthens whatever rules produced the result. Changes batch up and push to your stored spec when they hit a threshold.
Quick Start
1. Install the Skill
Download voice-fingerprint.skill from this repo. In Claude, go to Customize → Skills and upload the file.
2. Start the Interview
Open a new conversation and say:
"Learn my voice. I'm going to upload some writing samples."
Upload 3-8 samples. Mix formats — a blog post, an important email, a social post, a report. The more different they are, the better the skill maps your range.
3. Review the Analysis
Claude will show you what it found in your writing — patterns, tendencies, and conspicuous absences. It'll ask: "Here's what I'm hearing in your voice. Tell me where I'm right, where I'm wrong, and what I'm missing."
This is the calibration step. Push back where it's off.
4. Answer the Interview Questions
Claude will ask about deliberate choices, localization preferences, where you want variance, and what makes you cringe. Answer honestly — the spec is only as good as the input.
5. Approve the Spec
Claude generates a Voice Spec and asks you to review it. Once you approve, it saves a compact version to memory (persistent across conversations) and gives you the full version as a downloadable file.
6. Write
From now on, Claude uses your spec as a constraint set on writing tasks. Ask it to draft an email, write a blog post, or punch up a pitch — it writes as you, not as a helpful AI assistant.
7. Correct and Improve
When something's off, say so. "I'd never phrase it that way." "Too formal." "Lose the semicolons." Corrections accumulate and feed back into the spec. You can also say "update my voice spec" to trigger a manual revision.
What's in a Voice Spec
The spec isn't a mood board. It's a constraint document with teeth:
- Five Signature Traits — ranked, non-negotiable. If Claude gets nothing else right, it gets these right.
- Structural Patterns — sentence length range, paragraph rhythm, opening and closing moves, transition habits.
- Vocabulary & Register — formality position, jargon policy, Latinate vs. Anglo-Saxon lean.
- Rhetorical Moves — argument structure, humor type and frequency, authority stance.
- Mechanics — English variant, Oxford comma, em-dash usage, quotation style.
- Audience Modes — how your voice shifts for different readers.
- Controlled Variance — where to introduce randomness and how much.
- Banned List — hard bans (never, with reasons), soft avoidances (only when earned), structural bans, and cliché families grouped by the reason they're banned.
- Anti-Patterns — what your voice is NOT. The mistakes a mimic would make.
- Example Rewrites — generic sentences rewritten in your voice, showing the before/after.
- Version History — every revision tracked with what changed and why.
The Learning Loop
Most style tools are set-and-forget. Voice Fingerprint is set-and-sharpen.
Corrections — When you edit Claude's output or say "that doesn't sound like me," the skill identifies the specific violation, generalizes the pattern, and classifies it: new rule, tighter existing rule, or a contradiction that needs resolving. It proposes the update, you confirm, it writes.
Approvals — When you say "perfect" or "nailed it," the skill identifies which rules produced the result and strengthens them. Rules that get praised three or more times get promoted toward the Signature Traits.
Batching — The skill doesn't update the spec after every correction. Changes accumulate until they hit a threshold (three corrections in the same dimension, a contradiction with an existing rule, or your explicit request). Then it shows you the diff and asks permission.
Versioning — Every revision gets a version number and a changelog entry. You can see exactly how your spec evolved.
What It Doesn't Do
- Ghostwrite without samples. The skill needs your writing to learn from. No samples, no spec.
- Replace your judgment. The spec is a starting point. You still approve, edit, and redirect.
- Work outside Claude. This is a Claude skill. It uses Claude's memory system and skill architecture.
- Copy another writer. It captures your voice. If you want to sound like someone else, that's a different tool.
Requirements
- A Claude Pro, Team, or Enterprise account (skills require a paid plan)
- 3-8 writing samples (any format Claude can read — text, PDF, Word, pasted inline)
- 15-20 minutes for the initial interview
File Structure
voice-fingerprint/
└── SKILL.md # The complete skill — analysis, interview, spec
# generation, memory storage, and learning loop
FAQ
How long does the initial setup take? 15-20 minutes. The analysis is fast. The interview is where you spend time — and it's worth it.
Can I update my spec later? Yes. Say "update my voice spec" in any conversation. You can also upload new writing samples to extend it.
What if my voice changes over time? That's what the Learning Loop handles. As you correct and approve, the spec evolves with you.
Does it work for brand voices, not just personal voice? Yes. The same process works — just adjust the interview to focus on brand positioning and audience rather than personal preference.
What if I write differently for different audiences? The spec supports multiple Audience Modes. The interview asks about this specifically.
Can I see what's stored in memory? Yes. Ask Claude "what do you remember about my voice?" and it'll show you the stored spec.
Can I delete the spec? Yes. Go to Settings → Memory and remove the Voice Spec entry, or ask Claude to forget it.
License
MIT. See LICENSE.
Contributing
See CONTRIBUTING.md.
Built by Marcus Nelson.