You finally got Claude sounding like you in one chat. The cadence was right, the words were yours, you stopped flinching when you read the output back. And then you opened a new conversation the next day and Claude was back to writing like a LinkedIn coach with a thesaurus. I know the feeling. The fix is a Claude Skill, and once you build one for your brand voice, that one specific problem becomes solved forever across every conversation, every Project, every team member you bring in.
Skills are the thing nobody is talking about loud enough yet, and they should be. Let me walk you through what they are, and then how to build the single most valuable one for your business: your voice.
What a Claude Skill actually is
Anthropic introduced Skills in 2025 as a way to package up reusable instructions, examples, and reference files that Claude can load on demand inside a conversation. Think of it as a kind of recipe folder. You write the recipe once. You name it. Then when you (or your team) need that specific output, Claude knows to pull that recipe in and follow it.
The mechanics: a Skill is a folder with a description file (usually a SKILL.md or skill.md) that tells Claude when and how to apply it, plus any supporting files you want it to reference. Claude reads the description, decides whether the Skill is relevant, and loads it into context when it is. You don't have to remember to invoke it. The platform does the matching for you.
This is a different thing from a Project. A Project is a workspace with persistent files and instructions for a specific scope of work (your business, a particular client, a particular product line). A Skill is portable. You can use the same Skill across multiple Projects, share it with teammates, or upgrade it once and have every workflow that depends on it improve automatically. Voice rules belong in a Skill. Knowledge documents belong in a Project. Both layers matter.
The exact way you create and manage Skills depends on whether you're using Claude.ai, Claude Desktop, or Claude Code, and Anthropic has been shipping updates to this regularly. So I'll keep the steps tool-agnostic and you can apply them in whichever surface you use.
Why brand voice is the first Skill to build
Most business owners try to build their first Skill around a workflow, like an email sequence generator or a sales page outliner. That's a mistake. Workflow Skills are useful, but they'll only ever be as good as the voice they output in. If the voice is generic, the workflow Skill produces generic copy 50 times faster. That's worse than slow generic copy. That's scale on a broken foundation.
Voice is the foundation. Build it first. Once you have a voice Skill that genuinely sounds like you, every subsequent Skill (your email Skill, your social Skill, your podcast outline Skill) inherits the voice automatically and you stop fighting your AI on the same problem 200 times.
Step 1: Extract your voice
You can't write a voice Skill from memory. You have to extract it from actual writing you've done that you're proud of. Here's the process I use, which usually takes me about 2 hours one afternoon.
Pull together 8 to 12 pieces of your own writing. Emails, social posts, sales pages, podcast intros, anything where the voice was unmistakably you and the response from your audience was strong. Paste them into a single document.
Now open a Claude conversation and ask Claude to act as a voice analyst. Prompt it with something close to this: "Below are 10 samples of my writing. Analyze them and extract a voice profile. I want: sentence rhythm patterns, recurring words I actually use, words I appear to deliberately avoid, formatting tendencies, common rhetorical moves, and 3 to 5 example sentences that capture the voice at its strongest. Be specific. Don't give me adjectives."
Claude will give you back a draft voice profile. Read it. Some of it will be accurate, some will be off. Edit it. The edited version becomes the heart of your Skill.
Step 2: Structure the Skill
A good voice Skill has roughly six sections:
A short description at the top that tells Claude when to apply it. Something like "Apply this Skill any time you are writing copy, social posts, emails, or any content that goes out to my audience." This is the part Claude reads to decide whether to load the Skill.
A "voice in one paragraph" summary so a new collaborator (or a new Claude conversation) can get oriented fast.
A "use these words" list. Specific words you actually use. Not adjectives like "warm" but the literal words like "agents," "lenders," "systems," "stack."
A "never use these words" list. The banned terms. For me that's the entire corporate vocabulary plus the AI tells: use, use, dive, dig in, work through, path, open up, solid. Put yours in writing.
A "patterns to avoid" section. The triplet pattern, the question-then-list-answer pattern, the suspiciously perfect transition, the "it's not X, it's Y" formulation. If you don't put these in writing, Claude will drift back to them every time.
A "examples" section with 4 to 8 short passages of your actual writing. These are the highest signal part of the Skill. Claude pattern-matches against examples more strongly than against rules.
Save it. Name it something obvious like "krista-voice" or "[yourname]-voice." Make sure the description field is clear enough that Claude will know when to trigger it.
Step 3: Test it (this is where most people stop too soon)
Spin up a fresh conversation. Don't give Claude any other context, just trigger the Skill or upload it depending on your tool. Then ask Claude to write a piece of content you'd normally write yourself.
Read the output out loud. Yes really. Out loud. If it sounds like you ranting in your kitchen, congrats. If it sounds like a thoughtful corporate cousin of you, the Skill isn't done yet.
Common gaps I see on first drafts: not enough banned words listed (the AI tells slip back in), examples that are too short to anchor rhythm, rules that contradict each other so Claude picks the safer/blander rule. Go fix the Skill, test again. Three rounds and you'll have something that holds.
Pattern interrupt: my first voice Skill took me four attempts before it sounded right. The fourth attempt was different because I stopped trying to describe my voice and just dumped in 6 full paragraphs of stuff I'd actually written. The examples did most the work. So if you're stuck, add more examples.
A real example of the difference
Same prompt, "Write a 150-word LinkedIn post about why most agents fail at consistent content."
Without the voice Skill, Claude wrote something starting with "In today's competitive real estate landscape, content consistency separates thriving agents from those struggling to grow." That sentence alone has three of my banned phrases.
With the voice Skill loaded, Claude wrote, "Most agents don't fail at content. They fail at the 90 seconds before the content, which is the part where they decide what they actually think. Then they sit in front of a phone with nothing to say and call it writer's block." That second one I'd actually post. The first one I'd delete.
That gap is the gap a voice Skill closes.
If you want the full voice extraction template I use with my coaching clients, plus the rest of the AI authority system that sits around it, that's all inside Level Up. I built it specifically for coaches and consultants who need their AI to sound like them, not like everyone else's AI.