Making AI Sound Like Me

The default AI voice is corporate cheerful — competent but forgettable. "Leverage synergies to optimize outcomes." Nobody talks like that. I don’t talk like that. But every AI tool defaults to it because that’s what the training data is full of.

I tried the obvious fix: I gave Claude a style guide. "Write like Jenn: warm, direct, empirics-forward." It produced slightly warmer corporate copy. Still not me.

What actually worked was showing, not telling. I gave Claude my journal entries — the real ones, where I’m processing a hard conversation or working through a career decision. I gave it my Plaud transcripts — how I actually explain things to colleagues in real time. After enough examples, it picked up the cadence: short declarative sentences that set up longer analytical ones. Concrete details before abstract claims. Vulnerability that’s evidence, not appeal.

But the bigger lever was the forbidden list. What Claude must never write: - No "leverage," "synergies," "optimize," "utilize" - No hedging ("it may be worth considering") - No sycophantic framing ("This amazing approach") - No emoji unless I explicitly ask - No "In conclusion" or "To summarize" - Lead with the answer, then explain — never bury the point

I tested this by asking ChatGPT to rewrite my journal entries "in my voice." It wrote product copy. Clean, professional, and completely missing the point. That’s the whole lesson — voice isn’t transferable by instruction. It’s transferred by example. You can’t describe cadence. You can only demonstrate it enough times that the model picks up the rhythm.

The practical takeaway: put 3-5 samples of your actual writing in CLAUDE.md or a referenced file. Then add 10 lines of "never write this." The negative constraints do more work than the positive examples.