Field Notes Voice Transplant
The Setup
I write a lot. Blog posts (field notes on hazelq.com), social content, emails, proposals, case studies. Sometimes I draft with AI assistance. The problem isn’t the ideas or the structure — AI is great at both. The problem is the sound. AI-generated writing sounds like AI-generated writing. It’s correct, comprehensive, and completely devoid of personality.
My voice has specific qualities. I use contractions always. Short punchy sentences mixed with longer ones. Direct, not hedge-y. A little irreverent. I say “shit” sometimes. I never say “delve” or “elevate.” I don’t start sentences with “Certainly” or “Absolutely.” If it sounds like a LinkedIn post from 2019, it needs a rewrite.
The Build
I built a Claude Cowork skill that does voice transplants. It takes a draft — AI-generated, rough, or just not-quite-right — and rewrites it in my actual voice. The substance stays. The structure stays. The sound changes.
The skill has my full voice guide loaded: banned words, preferred patterns, tone-by-context rules (proposals are direct, Threads posts are casual, teaching content is encouraging). It reads the draft, identifies where the voice drifts (too corporate, too passive, too generic), and rewrites those sections while leaving the strong parts alone.
It’s not “make this better.” It’s “make this sound like me.” Those are very different instructions.
The Mess
The first version was too aggressive. It rewrote everything — even sentences that were already fine. The output had my voice but it lost the original thinking. Good ideas got paraphrased into mush because the skill was optimizing for voice when it should have been preserving substance.
The fix was a two-pass approach. First pass: identify which sentences need voice work and which are already fine. Second pass: rewrite only the flagged sentences. The result keeps the original structure and ideas intact while fixing the places that sound robotic.
Calibrating “how Hazel sounds” was also harder than expected. My writing voice changes by context. Proposals are confident and specific. Threads posts are warm and observational. Teaching content is encouraging but direct. The skill needed to know which mode to apply based on what it was rewriting.
The Result
I publish faster. First drafts come from AI or rough writing sessions. The voice transplant makes them sound like me in one pass. The editing step that used to take 45 minutes — reading every sentence and asking “would I actually say this?” — now takes 10 minutes of reviewing the transplant’s output.
The skill has also made my voice more consistent across formats. Whether it’s a blog post, a proposal, or a case study, it all sounds like the same person wrote it. Because the same voice rules apply everywhere.
The Takeaway
AI can write. It can’t sound like you. The skill that bridges that gap is the one that makes AI actually useful for creators who have a voice worth keeping.
How It’s Built
- Platform: Claude Cowork (custom skill)
- Input: Any draft (AI-generated, rough, or polished-but-wrong-voice)
- Voice rules: Banned words, preferred patterns, tone-by-context, examples
- Process: Two-pass (identify drift → rewrite flagged sections)
- Output: Same substance, same structure, Hazel’s voice
- Cost: Included in Claude subscription