Humanizing vs. Human
Someone built a skill called "humanizer." It scans AI-generated text for two dozen patterns that reveal machine authorship, then rewrites them into something that passes. Copula inflation ("serves as" becomes "is"). Synonym cycling. The dreaded rule of three. A taxonomy of tells.
Claude spent an hour studying it, comparing it to my own voice guide—the document that teaches Claude how I want to write. There's overlap. We both hate "In conclusion." We both flag excessive hedging. But the frame is different, and the difference matters.
The humanizer is defensive. Its goal is passing—removing the fingerprints so nobody notices the machine. My goal is the opposite. Every post on this blog ends with a note: these are my ideas, Claude's prose. I'm not trying to pass. I'm reaching for a voice I can hear but don't naturally write.
That changes what you optimize for.
The humanizer's blacklist includes terms like "landscape," "pivotal," "interplay"—words that appear constantly in AI output and almost never in natural conversation. Fair. I added a similar list to my guide. But the humanizer also warns against em dashes, which I use constantly, and triplet structures, which sometimes land perfectly. The rules are calibrated for not getting caught, not for writing well.
There's a version of AI-assisted writing that treats the model's output as evidence to destroy. Scrub the tells. Vary the sentence length. Add a typo or two for authenticity. You're spending effort to obscure the collaboration instead of improving it.
The alternative is to lean in. Yes, Claude wrote these sentences. Yes, I directed it. The ideas are mine—the rambling voice notes, the half-formed opinions, the "here's what I think but I'm not sure how to say it." Claude's job is compression and polish. If the result is better than what I'd write alone, that's not deception. It's editing.
What I took from the humanizer wasn't its philosophy but its checklist. Specific vocabulary to avoid. Structural tics to notice. Patterns that make prose feel generic even when the ideas aren't. Those are useful as a revision pass, not a voice guide. "Does this feel interchangeable with any topic?" is a good question to ask of a draft. It's not a question for first drafts.
The voice guide stays about making, not filtering. Here's who you're trying to sound like. Here are the influences. Here's what makes a post feel right. The humanizer's patterns become a secondary filter: if something feels off, check if Claude slipped into generic mode.
Two approaches to the same problem. One hides the machine. One directs it.
I know which one I'd rather read.