What you will build
A custom Claude skill containing your extracted voice profile: tone, sentence rhythm, vocabulary, signature moves, and a banned list, backed by real samples of your writing. Once enabled, it fires on any writing task and pulls every draft toward how you actually sound.
Why extraction beats description: asking yourself "what is my voice" produces vague adjectives. Analyzing 30 things you already wrote produces patterns you did not know you had. The samples are the source of truth, not your self-image.
Collect your writing samples
Target 20 to 50 pieces. Quality of sourcing matters more than volume.
- a.Emails: open your sent folder and pull 10 to 15 emails where you were explaining, persuading, or updating someone. Skip one-line replies.
- b.Blogs and newsletters: pull 5 to 10 pieces, including at least two you consider your best.
- c.Social posts: pull 10 to 20 captions or posts, favoring the ones that got real replies, not just likes.
- d.Bonus, voice notes: transcribe 2 or 3 voice notes or video transcripts. Spoken language reveals rhythm that polished writing hides.
- e.Paste everything into one document, one sample per section, labeled by source and date.
Only include writing that came out of your own head. Anything heavily edited by a team member, ghostwritten, or drafted by AI contaminates the extraction with someone else's patterns. Strip out client names and confidential details before the next step.
Sort samples by context
Your voice is not one thing. It shifts register between a client email and an Instagram caption, and the skill needs to know both the constant core and the per-context shifts.
- a.Group the document into three sections: Email, Long-form, and Social.
- b.Inside each group, mark your 2 or 3 favorites. These become the gold-standard reference samples in Step 6.
- c.Delete any sample that makes you cringe. The skill should learn from you at your best, not your rushed Tuesday.
Run the voice extraction analysis
Open a fresh Claude conversation, attach or paste the sample document, and run this prompt:
Analyze all attached writing samples and extract my voice profile.
Base every finding on evidence: quote at least two samples per finding.
Do not flatter me and do not generalize.
Extract:
1. TONE: how I come across, described in plain words with quoted proof
2. SENTENCE PATTERNS: typical length, rhythm, how I open and close,
how often I use fragments or questions
3. VOCABULARY: words and phrases I reuse, my go-to transitions,
how technical or casual my word choice runs per context
4. SIGNATURE MOVES: recurring structures, like how I open a post,
how I hand out advice, how I use examples or numbers
5. CONTEXT SHIFTS: what changes between my emails, long-form,
and social writing, and what stays constant everywhere
6. WHAT I NEVER DO: patterns absent from every sample, including
punctuation habits, structures, and phrase types I avoid
Format the output as a voice profile document with these six
sections. Under each finding, include the supporting quotes.- a.Read the output next to your samples and check that every claim has real quotes behind it.
- b.Push back on anything generic: "warm and authentic" describes half the internet. Ask Claude to replace vague findings with specific, quotable patterns.
Review and correct the profile
The extraction is a draft. You are the editor, and this pass is where the skill gets its accuracy.
- a.Delete findings that are technically present in the samples but not really you, such as a phase you have outgrown.
- b.Add what the analysis missed. You know things no sample shows, like topics you refuse to write about or the reason behind a habit.
- c.Rewrite every finding as a command. "She often opens with the reader's friction" becomes "Open with the reader's friction or behavior, then introduce the solution."
- d.Save the corrected version as voice-profile.md.
Build the banned list
The banned list does more work than the rest of the skill combined, because AI drafts fail on what they add, not what they miss.
- a.Start with section 6 of the extraction, the things you never do.
- b.Add the AI patterns you keep deleting from drafts: hype adjectives, contrasting fragments, rule-of-three groupings, "Picture this" openers, closers that summarize what was just said.
- c.Add punctuation rules, for example banning em dashes if you never use them.
- d.For each banned pattern, write what to do instead. A ban without a replacement just moves the problem.
Keep a running note on your phone. Every time you delete a phrase from an AI draft over the next two weeks, add it to the banned list. The list matures fast this way.
Write the SKILL.md and package the folder
Create the skill folder with this structure:
my-voice-skill/
SKILL.md
references/
voice-profile.md (the corrected profile from Step 4)
banned-list.md (the full list from Step 5)
samples-email.md (2-3 gold-standard emails)
samples-longform.md (2-3 gold-standard blog pieces)
samples-social.md (2-3 gold-standard posts)Then write the SKILL.md:
---
name: My Writing Voice
description: Apply my personal writing voice to all written content,
including emails, blog posts, newsletters, captions, social posts,
and scripts. Trigger on any drafting, writing, or rewriting task.
---
# My Writing Voice
## Core rules (apply everywhere)
- [paste the 5-8 strongest commands from voice-profile.md]
## Per-context adjustments
- Email: [register shift, length, sign-off habits]
- Long-form: [structure, pacing, how examples are used]
- Social: [hook style, line breaks, CTA habits]
## Process
1. Read references/voice-profile.md before drafting
2. Match the draft type to the right samples file and read it
3. Draft the piece
4. Check the draft against references/banned-list.md
5. Rewrite any sentence that violates a rule before presenting
## Never
- [top 5 banned patterns, stated as commands]Upload and enable the skill
- a.Zip the skill folder so the folder itself is the root of the archive.
- b.In Claude, open Customize, go to Skills, click +, then + Create skill, and upload the zip.
- c.Toggle the skill on. It now applies in chats and syncs into Cowork sessions, including scheduled ones.
Sample files travel with the skill. Double-check that every email sample has client names, financial details, and anything confidential removed before you zip.
Blind test against your real writing
Do not judge the skill by whether the output "feels right" to you in the moment. Test it properly.
- a.In a fresh conversation, generate three pieces: one email, one social post, one longer piece, without naming the skill in your prompt.
- b.Mix them with three real pieces you wrote, and ask someone who reads your content regularly to spot the fakes.
- c.For every piece they catch, ask what gave it away. That answer is a missing rule. Add it to the profile or the banned list.
- d.Re-zip, re-upload, retest. Two or three rounds usually closes the gap.
- e.Refresh the samples every few months. Your voice evolves, and the skill should evolve with it.
FAQs
How many samples do I really need?
20 to 50, across at least two contexts. Under 20, the analysis latches onto one-off quirks instead of patterns. Past 50, you get diminishing returns and a slower extraction.
What if I do not like my own voice?
Extract from your favorites only. The Step 2 filter, keeping the pieces you are proud of and deleting the ones you cringe at, means the skill learns your best register, which is still authentically yours.
One voice skill or several?
One skill with per-context adjustments works for most founders. Split into separate skills only if your registers genuinely conflict, for example a formal B2B email voice and a playful creator voice.
Can this skill run alongside my design or repurposing skills?
Yes. Skills compose, so the voice skill handles how things are written while a design skill handles how pages look and a repurposing skill handles the workflow. Keep each one focused on its own layer.
Troubleshooting
| Symptom | Likely cause | Fix |
|---|---|---|
| Output still sounds like AI | Banned list too short or vague | Add the exact phrases you keep deleting, each with a replacement behavior |
| Output sounds like a caricature of you | Signature moves overweighted | Add frequency limits: "use this opener at most once per piece" |
| Voice right in posts, wrong in emails | Missing per-context adjustments | Expand the context section and add email gold-standard samples |
| Skill does not fire on writing tasks | Description too narrow | List every content type in the description: emails, captions, scripts, posts |
| Profile findings feel generic | Samples too few or too similar | Add more samples from a second context and rerun the extraction |
Version 1.0 · First published July 2026