How to Build a Voice Profile for Your AI Agent: The Claude Co-work Interview Method
Use Claude Co-work's voice profile skill to extract your writing patterns from LinkedIn posts and emails and generate a reusable voice profile markdown file.
Why Your AI Agent Sounds Generic (And How to Fix It)
If you’ve ever read AI-generated content and immediately thought “that doesn’t sound like me,” you’ve already identified the core problem with most AI writing setups. The model isn’t broken — it just doesn’t know you.
Voice profiles solve this. A voice profile is a structured document that captures how you write: your sentence rhythm, word choices, tonal defaults, things you never say, and the subtle patterns that make your content recognizable. When you feed a voice profile into your AI agent as context, the output stops sounding like a language model and starts sounding like a person.
The Claude Co-work Interview Method is a systematic way to extract that profile from real writing samples — your LinkedIn posts, emails, newsletters, whatever you actually produce — and package it into a reusable markdown file your agents can reference every time they generate content on your behalf.
This guide walks through the full process, from gathering samples to producing a finished voice profile you can drop into any workflow.
What a Voice Profile Actually Is
A voice profile isn’t a style guide. Style guides tell you what to do. A voice profile describes what you already do.
It’s a reference document — typically 500 to 1,500 words — that captures:
- Sentence structure patterns: Do you write short declarative sentences or longer flowing ones? Do you use fragments for effect?
- Vocabulary preferences: Formal vs. casual registers, industry terms you favor or avoid, filler words you habitually use
- Punctuation habits: Em dashes, Oxford commas, ellipses — these are fingerprints
- Structural instincts: Do your posts open with a hook question? Do your emails lead with context before getting to the ask?
- Tone defaults: Are you direct or warm? Sardonic or earnest? Both depending on context?
- Things you never do: This is often as important as what you do
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
A well-built voice profile makes the difference between an AI agent that sounds like it was trained on everything and one that sounds like it was trained on you.
Why the Interview Method Works Better Than Describing Yourself
The obvious approach is to just tell Claude what your voice is like. This almost never works well.
Most people are bad at describing their own writing. You know what you mean to do, but your actual patterns are often different. You might think you write conversationally, but your emails are actually quite formal. You might believe you’re concise, but your LinkedIn posts average three paragraphs before reaching the point.
The interview method sidesteps self-perception by working from evidence. You feed in real writing samples, and then Claude asks you targeted questions about those samples — surfacing patterns you didn’t consciously notice, asking you to confirm or reject observations, and helping you articulate the “why” behind stylistic choices.
The result is a profile grounded in what you actually produce, not what you think you produce.
This is particularly important for building AI agents that generate content at scale. The more accurately the profile reflects reality, the less editing you’ll need to do on outputs.
Step 1: Gather Your Writing Samples
Before you start the interview, you need raw material. The goal is to collect 10 to 20 pieces of writing that represent your normal output — not your best work, your typical work.
What to Collect
LinkedIn posts work well because they’re relatively short, they’re public-facing (so you were being intentional), and they cover a range of topics. Aim for 8 to 12 posts across different subjects.
Emails are useful because they reveal your natural voice under low-stakes conditions. Pull from threads where you were writing normally, not formal client correspondence or legal communications. Direct messages to colleagues are ideal.
Newsletter sections, blog posts, or Substack pieces if you have them — these tend to show your longer-form patterns.
Slack or Discord messages can be surprisingly useful for capturing your informal register.
What to Avoid
Don’t include:
- Writing that was heavily edited by someone else
- Extremely formal documents (contracts, press releases)
- Very old samples that don’t reflect how you write now
- Content you’re not proud of — but only because it tends to reflect situations where you weren’t fully yourself, not because you need “best of” material
How to Format Them
Copy the raw text into a single document or keep them as separate labeled blocks. The labels matter — you’ll want to reference them during the interview. Something like:
[SAMPLE 1 - LinkedIn post, March 2024]
Your post text here...
[SAMPLE 2 - Email to team, informal update]
Your email text here...Step 2: Set Up the Claude Co-work Interview Session
The Co-work Interview Method uses Claude as an analytical partner, not just a generator. You’re asking it to observe, analyze, and ask you questions rather than produce content.
The Initial System Prompt
Start a new Claude conversation with this framing:
You are helping me build a voice profile for use in AI content generation.
I'm going to share writing samples I've produced. Your job is to:
1. Read all the samples carefully before commenting
2. Identify patterns in sentence structure, vocabulary, tone, rhythm, and formatting
3. Ask me clarifying questions about ambiguous or interesting patterns you notice
4. Wait for my answers before drawing conclusions
5. At the end, synthesize everything into a voice profile markdown document
Do not generate any content in my voice yet. Only analyze and interview.Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
This framing is important. Without it, Claude tends to jump to synthesis too quickly, before it’s surfaced the nuances worth capturing.
Paste Your Samples
Share all your writing samples in a single message, labeled as described above. Add a note at the end:
Please read all of these samples and then share your initial observations
about my writing patterns. Then ask me your first set of questions.Step 3: Run the Interview
This is the core of the method. The interview typically takes 20 to 40 minutes and unfolds in rounds.
Round 1: Pattern Identification
Claude will identify what it observes: things like “you frequently open with a direct question,” “your sentences tend to run 10 to 15 words,” “you use em dashes significantly more than standard writing,” or “you avoid passive voice almost entirely.”
Read these observations carefully. Some will be accurate. Some will be partially right. Some will miss the mark.
Respond to each observation. Confirm what’s accurate, correct what’s wrong, and — most importantly — add context. If Claude notices you use humor often, explain when and why. If it flags that your emails are short, explain whether that’s intentional or situational.
Round 2: Edge Cases and Exceptions
After the initial exchange, Claude should ask about exceptions and context-dependence. Good questions at this stage look like:
- “Your LinkedIn posts use a lot of rhetorical questions, but your emails don’t. Is that intentional?”
- “You seem to vary your formality significantly across samples. What’s driving that?”
- “I noticed you never use bullet points in your emails but use them frequently in posts. Is that a formatting rule?”
Answer these as honestly and specifically as you can. The goal is to give Claude enough signal to build rules, not just observations.
Round 3: Things You Never Do
This is often the most valuable round. Ask Claude explicitly:
Based on my samples, what are the stylistic choices that I consistently
avoid? What patterns do you NOT see that other writers typically use?Common answers include things like: “You never use motivational language or phrases like ‘excited to share’”; “You don’t use numbered lists in long-form writing”; “You avoid adjective stacking”; “You don’t use exclamation points except in casual messages.”
These negative patterns are essential. When you’re generating content at volume, your AI agent’s default tendency will be toward average behavior. The “never do” list is what keeps it from drifting.
Step 4: Generate the Voice Profile Markdown File
Once you’ve completed two to three rounds of interview, ask Claude to synthesize:
Based on everything we've discussed, please generate a comprehensive voice
profile in markdown format. Structure it with these sections:
- Writing style overview (3-5 sentences)
- Sentence and paragraph patterns
- Vocabulary and word choice
- Tone and register
- Structural habits (how I open, develop, and close content)
- Formatting preferences
- What I never do / avoid
- Examples of phrases or constructions I use often
- Instructions for any AI model using this profile
Make it specific enough that an AI model with no prior knowledge of me
could generate content that sounds like I wrote it.Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
What a Good Voice Profile Looks Like
A finished voice profile might include sections like:
## Voice Profile: [Your Name]
### Style Overview
Direct and information-dense. Prefers short sentences with a clear subject
and verb. Occasionally uses fragments for emphasis. Assumes a literate
audience that doesn't need hand-holding.
### Sentence Patterns
- Typically 8–15 words per sentence
- Opens paragraphs with declarative statements, not questions
- Uses em dashes to insert parenthetical thoughts rather than parentheses
- Occasional one-sentence paragraphs for emphasis
### What I Never Do
- No motivational language ("excited to share," "thrilled to announce")
- No bullet lists in emails — only in posts or docs
- No passive voice in the first sentence of any paragraph
- No exclamation points in professional contentSave this as a .md file. This is your reusable artifact.
Step 5: Test and Refine the Profile
Before deploying the profile in any workflow, test it. Take it into a new Claude conversation with no prior context and paste it as a system prompt. Then ask Claude to write something typical for you — a LinkedIn post about a recent professional experience, or a reply to a common type of email.
Read the output and ask: does this sound like me?
If the answer is “close but not quite,” go back into the interview and identify what’s missing. Common gaps at this stage:
- Topic-specific voice shifts: You might write very differently about technical topics vs. leadership topics. You may need topic-specific addenda to the profile.
- Audience-specific registers: Your voice with prospects might differ from your voice with direct reports.
- Cultural references and examples: If you habitually draw from a particular domain (sports, film, history), that’s worth capturing.
A second or third iteration of the interview process dramatically improves accuracy. Most people find the profile is 70% accurate after the first round and 90%+ accurate after refinements.
How to Use Your Voice Profile in AI Workflows with MindStudio
A voice profile sitting in a markdown file isn’t doing much work. The value comes from deploying it consistently across the AI agents that generate content on your behalf.
MindStudio makes this straightforward. You can build AI agents that automatically load your voice profile as part of their system context — meaning every piece of content they generate already has your voice baked in before you’ve written a single instruction.
Here’s a practical setup: build a LinkedIn post generator agent in MindStudio that accepts a topic or talking point as input, loads your voice profile markdown file as a system-level instruction, and returns a ready-to-post draft. The whole build takes under 30 minutes using MindStudio’s visual builder, and once it’s live, you’re generating on-brand content without thinking about tone on every run.
You can extend this further with automated content workflows that chain generation tasks together — drafting, reviewing for voice consistency, formatting for different platforms — all informed by the same profile.
MindStudio supports 200+ AI models, so you can run your voice profile through Claude, GPT-4, or any other model you prefer without rebuilding your setup. The profile is model-agnostic; it’s just context.
You can try it free at mindstudio.ai.
Common Mistakes to Avoid
Making the Profile Too Abstract
“I write in a conversational but professional tone” is nearly useless as instruction. Specificity is the entire point. Replace abstract adjectives with observed behaviors: “I use contractions in all writing except formal proposals” is actionable. “Conversational” is not.
Including Aspirational Voice
Don’t describe the writer you want to be. Describe the writer you are. Your AI agent will use this profile to match your existing output — if you describe an idealized version, the outputs will feel off and you’ll spend time editing toward your actual voice anyway.
Skipping the “Never Do” Section
Most people building voice profiles focus on what they do. The negative space — what you consistently avoid — is often more distinctive. Generic AI outputs tend toward the average; the “never do” list is what prevents your agent from sounding average.
Treating the Profile as Permanent
Your writing evolves. Revisit your voice profile every six months or so. Run the interview process again with fresh samples. Update the file. Agents using outdated profiles will gradually drift from your current voice.
Frequently Asked Questions
What is a voice profile for an AI agent?
A voice profile is a structured reference document that describes how a specific person writes — their sentence patterns, vocabulary preferences, tone, structural habits, and stylistic rules. When used as context in an AI agent’s system prompt, it guides the model to generate content that sounds like that person rather than like a generic language model output.
How many writing samples do I need to build an accurate voice profile?
Ten to twenty samples is a practical minimum for useful pattern detection. More samples generally produce more accurate profiles, especially if they span different content types (posts, emails, long-form writing). Below ten samples, Claude may pick up on patterns that are idiosyncratic to individual pieces rather than consistent habits.
Can I use a voice profile across different AI models?
Yes. A well-written voice profile is model-agnostic — it’s plain text context that any language model can interpret. You may notice slight differences in how different models apply the same profile, but the core output should be consistent. Test your profile across whichever models you plan to use and adjust wording if needed.
How do I keep my voice profile up to date?
Treat your voice profile like a living document. Revisit it every six months or when your writing style shifts significantly — after a career change, a new audience, or a deliberate style evolution. Run a fresh interview session with recent samples rather than trying to manually edit the existing profile.
What’s the difference between a voice profile and a brand style guide?
A brand style guide defines rules for how an organization communicates — approved terminology, tone standards, formatting requirements. A voice profile captures how a specific individual actually writes, based on observed patterns. They serve different purposes: style guides are prescriptive and institutional; voice profiles are descriptive and personal.
Can I build separate voice profiles for different contexts?
Yes, and for many people this is the right approach. Your voice in thought leadership posts may differ significantly from your voice in sales emails or internal communications. You can build a core profile and then add context-specific addenda, or maintain entirely separate profiles for different use cases. Label them clearly so your AI agents load the right one.
Key Takeaways
- A voice profile captures your actual writing patterns — not your ideal ones — and makes them reusable across AI workflows.
- The Claude Co-work Interview Method works by analyzing real samples rather than relying on self-description, which produces more accurate profiles.
- The interview runs in rounds: pattern identification, edge case clarification, and negative pattern mapping (“what I never do”).
- The output is a markdown file you can drop into any AI agent as persistent system context.
- Testing against real writing tasks is essential — most profiles need one or two refinement rounds to reach high accuracy.
- Tools like MindStudio let you operationalize your voice profile by embedding it into automated content workflows, so every agent output starts from the right foundation.
Building a voice profile is a one-time investment that pays dividends across every piece of content your AI agents produce. Once it’s in place, you spend less time editing and more time on the work that actually requires your judgment.
