Teaching Claude Your Voice (Without Just Describing It)

I typed "conversational but professional" into a prompt and stared at it for a second before hitting send. It's the phrase everyone uses to describe their voice. It's also almost meaningless. Ten different writers could read that phrase and produce ten different voices, none of which would sound like me.

That was the first attempt at rebuilding my social-media-writer skill: describe the voice, hope Claude fills in the gaps correctly. It didn't work. The output was fine. Competent, even. It just didn't sound like anything I'd actually write.

So I changed the approach. Instead of describing my voice, I pulled real transcripts from Cowork sessions, the actual unedited way I talk when I'm working through a problem out loud, not the polished version I post. I pasted those in and asked Claude to tell me what it noticed.

Here's what came back, and it was more specific than anything I could have described myself: I never use em dashes, a rule I'd stated explicitly in a past session and apparently follow without thinking about it anymore. I pair every decision with the reasoning behind it, I don't just state a conclusion. I acknowledge something before I pivot away from it, rather than jumping straight to the new point.

Claude pulled two actual examples from my own writing to use as calibration anchors, real sentences it could point back to and say "this, do more of this."

That's the pivot. Describing your voice makes you translate something instinctive into adjectives, and adjectives are lossy. Showing your voice skips the translation step entirely. Claude wasn't guessing what "conversational but professional" might mean. It was pattern-matching against something real.

The social-media-writer skill got rewritten from the ground up based on that analysis, and it's the version I still use. It's live now, so if you want the exact skill instead of just the idea, you can get it directly.

Here's the actual lesson: when AI output doesn't sound like you, the fix usually isn't the AI. It's that you described your voice instead of showing it. Adjectives are a compression of something that already exists in your actual writing, and compression loses the specific texture that makes a voice recognizable.

The prompt

Paste 3-5 real, unedited samples of how you actually communicate — Slack
messages, voice memo transcripts, texts, whatever isn't polished. Then:

Analyze these samples of how I naturally communicate. Identify:
1. Sentence rhythm — typical length, how I open and close a thought
2. Words or phrases I reach for repeatedly
3. Formatting habits (what I avoid, how I structure lists or emphasis)
4. How I build an explanation or argument
5. Two passages from these samples that best represent my voice — use
 them as calibration anchors for anything you write for me going forward

If you're trying to get AI to write like you, skip the adjectives. Paste in a real transcript, an unedited email, a voice memo you transcribed, anything that's actually you talking without the editing pass. Ask what it notices before asking it to write anything. The patterns it finds will be more accurate, and more useful, than any description you could come up with on your own.