How I Built a Tone of Voice Skill for Claude Code

AI content sounds like AI. You can spot it in two sentences. The vocabulary is off, the structure is too clean, and it reads like a press release instead of a person talking. I built a Claude Code skill that fixes this by extracting my actual voice from real transcripts and turning it into a reusable reference layer.

This post covers the full process. How I collected the raw material, how Claude analyzed it, and how the final Claude Code skill works in practice. I also run a live before/after demo so you can see the difference.

Step 1: Collect Raw Transcripts from Different Settings

I gathered 7 transcripts that captured me speaking in different contexts:

The key: multiple settings. I talk differently to my business partner than I do to senior managers. The Claude Code skill needed to capture the full range, not one mode.

Step 2: Analyze Each Transcript for Patterns

Claude read each transcript and extracted:

• Sentence structure: length, fragments, how ideas connect
• Vocabulary: go-to intensifiers, qualifiers, connectors, filler words
• Opening patterns: how I start a thought
• Closing patterns: how I end one
• Reasoning style: how I build an argument
• Emotional expression: how I show enthusiasm, frustration, curiosity
• Conversational habits: repetition, confirmation checks, analogies

Step 3: Find the Universal Voice vs. Situational Adjustments

The early versions tried to create separate voice profiles per register (formal, casual, vendor, etc.). We scrapped that.

The insight: one unified voice with minor situational adjustments. The core patterns (short sentences, "I think" before opinions, "super" as intensifier, practical analogies) stay constant. Only small things shift:

Step 4: Build the Claude Code Skill as a Reference Layer

This is where most people get it wrong. They try to build a skill that writes content. That does not work. The Claude Code skill I built is a voice reference that other skills load. It does not create content. It controls how content sounds.

Structure of the skill:

• Sentence structure rules (do/don't with before/after examples)
• Vocabulary (signature phrases, words to avoid, replacement table)
• Opening patterns (what to use, what to avoid)
• Closing patterns (same)
• Reasoning and argument style
• Emotional expression per emotion
• Thought flow: situation > opinion > rationale > compare > ask
• Conversational habits

Step 5: Create a Polish Level System

Instead of one tone, the Claude Code skill has 5 levels. All use the same voice at different levels of cleanup:

Step 6: Wire It into Other Claude Code Skills

The tone of voice skill gets loaded by my content skill, personal brand skill, social post skill, hooks skill, and writing skill. Each of those handles format (LinkedIn post structure, email structure, etc.). The tone skill handles how it sounds.

So when I ask Claude Code to write a LinkedIn post, the content skill handles the format and the tone skill makes sure it sounds like me. Two layers, one output.

Step 7: Build an Audit Mode

I can paste any piece of content and run a voice audit. The skill returns:

• Overall voice match (High / Medium / Low)
• Number of issues found
• Full corrected version with inline [VOICE FIX: old -> new] notes
• Top 3 voice mismatches

This is where the before/after demo comes in. Let me show you what it looks like in practice.

Live Demo: Before and After

I asked Claude to write a general paragraph about automation. No tone of voice skill loaded. Here is what came out:

BEFORE (generic AI output)
─────────────────────────
In today's rapidly evolving digital landscape, it's worth noting
that leveraging automation tools has become increasingly crucial
for businesses seeking to enhance their operational efficiency.
Furthermore, by utilizing comprehensive workflow platforms,
organizations can essentially streamline their processes,
ultimately empowering teams to focus on more strategic endeavors.
Prior to implementing these solutions, however, it would be
advisable to conduct a thorough assessment of your existing
infrastructure in order to ensure seamless integration.
The key takeaway here is that embracing automation isn't just
about saving time. It's about unlocking transformative potential.

Then I ran the voice audit on it:

VOICE AUDIT
───────────
Overall match: Low
Issues found: 14

CORRECTED VERSION
─────────────────
Automation saves you time.
[VOICE FIX: "In today's rapidly evolving digital landscape,
it's worth noting that leveraging automation tools has become
increasingly crucial" -> direct statement]

That's basically the whole point.
[VOICE FIX: "Furthermore, by utilizing comprehensive workflow
platforms, organizations can essentially streamline their
processes" -> simplified]

You set it up once, it runs every time after. 100%.
[VOICE FIX: removed em dash construction, added signature phrase]

Before you build anything, check what you already have.
[VOICE FIX: "Prior to implementing these solutions, however,
it would be advisable to conduct a thorough assessment of your
existing infrastructure in order to ensure seamless integration"
-> plain language]

Does your current stack support it? If not, why not?
[VOICE FIX: added signature accountability phrase]

I think most people overcomplicate this. Just automate the
repeatable stuff and move on.
[VOICE FIX: "embracing automation isn't just about saving time.
It's about unlocking transformative potential" -> opinion
framed with "I think", no fluff]

TOP VOICE MISMATCHES
────────────────────
1. "In today's rapidly evolving digital landscape"
   -> Open with the point. No preamble.
2. "leveraging", "crucial", "enhance", "utilizing",
   "streamline", "empowering", "endeavors", "seamless",
   "transformative", "unlocking"
   -> 10 forbidden words in 4 sentences
3. "The key takeaway here is"
   -> No summaries. End on the point.

And here is the clean final version after running through the tone of voice skill:

AFTER (tone of voice applied)
─────────────────────────────
Automation saves you time. That's basically the whole point.

You set it up once, it runs every time after. 100%.

Before you build anything, check what you already have.
Does your current stack support it? If not, why not?

I think most people overcomplicate this.
Just automate the repeatable stuff and move on.

Step 8: Add a Self-Check for Quality Control

Every Claude Code skill that loads the tone of voice runs 10 checks before returning output:

1. 1.No sentence longer than 25 words
2. 2.At least one "I think" or "I'm wondering" in opinion pieces
3. 3.No forbidden vocabulary (utilize, essentially, arguably, etc.)
4. 4.No formal transitions (Furthermore, Moreover, Additionally)
5. 5.No em dashes
6. 6.No inspirational sign-offs
7. 7.No summaries at the end
8. 8.Opens with the point
9. 9.Reads like someone talking
10. 10.At least one signature connector in conversational formats

A Note on How I Use This

This Claude Code skill is not meant for everything I write. It captures my base tone. How I actually talk. After that, it still gets adjusted for the specific channel.

A formal email to a senior stakeholder sounds different from a LinkedIn post. A Slack message to my team sounds different from a vendor email. That is normal. In real life, you adjust how you speak depending on who you are talking to, the channel, and the setting.

The skill gets the foundation right. It makes AI output sound like me instead of a generic language model. Channel-specific tailoring happens on top. I actually recommend doing this. Build your base voice first, then adjust per channel. Not the other way around.

The Core Idea

Your voice already exists. It is in your recordings, your transcripts, your dictated notes. The process is extraction, not invention. Feed real speech into an LLM, have it identify patterns across multiple contexts, then codify those patterns into a Claude Code skill that any content tool can load.

I spent a few hours building this. Every piece of content I produce now sounds like me instead of ChatGPT. That trade-off is worth it.