The Claude voice-training prompt stack

Sounding human became a priced skill

The reason this is worth thirty minutes is not vibes. Companies are paying enormous money for people who can make things sound like a person wrote them. Netflix listed a director of product and technology communications role paying up to $775,000 a year, OpenAI has communications openings around $400,000, and the same reporting puts the average US director of communications at about $107,000. Fortune covered the wider pattern of seven-figure packages for communications hires at AI companies. None of those postings ask you to code.

The title moving fastest is storyteller. The share of US LinkedIn postings using the word doubled in the year to late November, covering roughly 50,000 marketing listings and more than 20,000 under media and communications, and executives said "storyteller" or "storytelling" on earnings calls 469 times in 2025 against 359 across all of 2024.

Pick three things you actually wrote

The whole system is only as good as what you feed it, and three pieces is the minimum that lets Claude tell a habit apart from a one-off.

1. 📝 One piece you wrote carefully. A report, a long post, an essay, a cover letter you were proud of. This shows your structured voice, meaning how you build an argument and how formal you get when you are trying.
2. 💬 One piece you wrote fast. A long message to a colleague, a Slack explanation, a comment you left somewhere. Your real rhythm lives here, and this is the sample that stops the output reading like a template.
3. 🎯 One piece in the register you want more of. If the goal is better emails, feed it your best email. The profile leans towards whatever you give it most of, so choose on purpose.
4. 🚫 Nothing an AI touched. Feeding in something a model helped with teaches it to imitate the exact sound you are trying to escape. Three to five hundred words per sample is plenty.

Profile your natural voice

This one is long on purpose. A vague request gets you a profile that says you write "clearly and engagingly", which describes every human on earth and is therefore useless as an instruction.

I am pasting three things I wrote myself, with no AI involved anywhere in them. Reverse-engineer how I write and turn it into a voice profile you can follow from now on.

Do not compliment me, and do not describe my writing with adjectives like clear, engaging or conversational, because those describe nobody. Measure it instead. Give me my average sentence length and how much it swings, how I open and close a piece, how often I use contractions, questions and fragments, my punctuation habits including whether I use semicolons or dashes at all, my default level of formality, and the words and constructions I reach for repeatedly. Separately list the moves I never make, because the absences are as distinctive as the habits.

Then write the profile as direct instructions to you, in the second person, 300 to 500 words, using rules I could actually check. "Keep sentences short" is not checkable. "Average 14 words, drop to 4 words when landing a point" is. Finish with three lines quoted from my samples that are the most characteristically me, one sentence each on why, and a short do-not list of moves that would immediately read as not me.

Here are the samples:
[paste sample 1]
---
[paste sample 2]
---
[paste sample 3]

When it comes back, read it as a spec rather than a compliment. A profile is working when a stranger could follow it and produce something recognisably yours. These four things are what separate a usable profile from a horoscope.

Save the profile so it loads every time

A profile sitting in one chat is dead by Thursday. This prompt compresses it into something you can paste into a project, and into a skill file if you want it portable across every chat you have.

Compress the voice profile we built into two reusable formats.

Version one is a project instructions block, under 1,500 characters, that keeps every checkable rule and drops all the commentary and analysis. Lead with the three rules that matter most, because those are the ones you need to hold onto when a request gets long and complicated and you start drifting back to your default voice.

Version two is a SKILL.md file for a skill called my-voice. Give it a lowercase hyphenated name, a description line that states exactly when it should trigger (for example, use whenever you draft or edit any prose for me), and then the same rules written as instructions. Keep the whole file under one page. At the end, tell me how to save it as a zipped folder so I can upload it.

1. 🗂️ Create a project. On claude.ai or in the Claude app, create a project and call it something you will recognise, like "My writing". Projects work on every plan, and free accounts can have up to five, per Anthropic's projects documentation.
2. 🧭 Paste version one into the project instructions. That field is applied to every conversation inside the project, so your rules are loaded before you type a word.
3. 📎 Upload your three samples to project knowledge too. The instructions give Claude the rules, the samples let it hear the music. Both together is noticeably better than either alone.
4. ➕ Optional: install version two as a skill. In the Claude app, go to Customize, then Skills, then Add, then Upload a skill, and upload the zipped folder. A skill works across all your chats rather than one project. Anthropic's create custom skills article has the full walkthrough, and skills are available from Free through Enterprise.
5. ✍️ Then actually work inside it. This is the part people drop. If you draft in a random new chat out of muscle memory, you get default voice back and conclude the setup did not work.

Flag every line that reads like slop

This is the prompt I use most, and it works on anything: your own draft, a teammate's, or something a model produced. It is an edit pass, not a rewrite, so you stay the author.

Audit this draft against my voice profile and flag every line that reads like default AI writing. Here is the draft: [paste it].

Give me a table with four columns: the line as written, which rule from my profile it breaks, how serious it is (high, medium or low), and a rewrite in my voice. Only flag genuine breaks. If you flag everything I will stop reading the table, so hold the bar high and leave lines alone when they are fine.

Pay specific attention to runs of sentences that are all the same length, openings that announce what the piece is about instead of starting it, the "it is not X, it is Y" construction, tidy three-part lists I did not ask for, hedging that removes any actual position, and words nobody says out loud like delve, tapestry, landscape, testament, or leverage used as a verb.

Finish with one honest verdict sentence and the single change that would do the most to make this sound like a person wrote it.

Make it interview you before it drafts

This is the beat almost everyone skips, and it is the one that keeps the ideas yours. A model asked to write about a topic writes the average of everything it has read. A model that interviews you first has to build the piece out of your material, which is the only material nobody else has.

Before you draft anything, interview me. I am writing [what it is] for [who it is for], and what I want them to think or do afterwards is [the outcome].

Ask me one question at a time and wait for my answer before asking the next. Go after the things only I have: the specific thing that happened, the number or detail from my own work, what I actually believe about this including the part that might be unpopular, and how I would explain it to a friend over coffee. If I answer in generalities, say so and ask the same thing in a narrower way, because the vague answer is exactly where the writing goes generic.

Stop after about eight questions and show me the raw material you collected as a bullet list so I can correct it and add anything you missed. Only then write the draft, following my voice profile, and build it out of my answers rather than general knowledge about the topic. Mark any sentence where you filled a gap with something I did not tell you, so I can check it or cut it.

How the four fit together each week

1. 🎤 Start with the interview. Open your project, run prompt four, answer honestly, and let it draft from your answers. Two minutes of answering beats twenty minutes of re-prompting.
2. 🔎 Run the slop audit on the draft. Accept the high-severity rewrites, argue with the rest. Every argument you win is a rule that needs sharpening in the profile.
3. 🌱 Do the last mile yourself. Add the example only you have, cut the paragraph that restates the previous one, read it out loud, and fix anything you stumble over. Put one clear opinion in it, because hedging is what makes writing forgettable.
4. 🔁 Refresh the profile monthly. Paste in the two or three things you wrote most recently and ask what changed. Your voice moves, and a profile from last year quietly makes you sound like an older version of yourself.

Four places this turns into career proof

A voice profile in a project is a nice tool. Pointed at the right surfaces, it becomes evidence you are the person who makes things sound human, which is the thing being paid for.

Where the numbers come from

• Netflix at $775,000, OpenAI around $400,000, average comms director about $107,000: Entrepreneur
• Seven-figure communications packages at AI companies: Fortune
• Storyteller postings doubling, and the earnings-call counts: summary of the Wall Street Journal report
• Claude projects and plan availability: Anthropic support
• Creating and uploading custom skills: Anthropic support
• The sourced list of AI writing tells: Wikipedia, Signs of AI writing