How to write better with AI (and not sound like a robot)
I put this together for the new AI users on my team, to help them make their written AI output sound like them. And I can't decide which frustrates me more: that people are too lazy to edit their AI outputs, or that Anthropic and OpenAI haven't devoted one of their big brains to stamping out the tells at the source.
Readers have learned to smell AI (and if you use AI all day long, you ID the tells in about a second). The moment they catch it, something worse than boredom happens: they stop hearing you and you lose credibility. If you write with AI, and you should, the job isn't hiding the tool. The job is making sure what ships sounds like you and is true.
The tells have moved.
Everyone knows the word list. Delve. Robust. Seamless. Game-changer. That list is a moving target: it shifts with every model release, and plenty of flagged words are normal vocabulary in the right room. "Leverage" is doing honest work in a DSCR memo. One buzzword is a word. Five on a page is a tell.
The durable AI tells are structural:
"It's not just X, it's Y."
Ideas that arrive in matched sets of three.
An em dash in every paragraph.
Bullets so uniform they look extruded.
An opening that flatters and a closing that recaps.
Hedges that protect the writer: "it's important to note."
Enthusiasm with no variance. When everything is exciting, nothing is.
Frequency is the tell, not the instance.
The worst tell isn't a word, it's a number.
Tone that's off embarrasses you. A specific that's wrong costs you. The model will state a fee or a close time with total confidence, and sometimes it's inventing it. In our business a transposed number isn't a style problem, it's a compliance problem. Nothing else in this post matters if the specifics don't get checked.
Constrain the draft up front.
Most people prompt for content and edit for voice. It's cheaper to demand the voice before the draft exists. Here are the instructions that I find do the most work:
Voice. "Write it the way I'd say it out loud. Short sentences. Plain words."
Examples. "Here are three things I actually wrote. Match this voice." Samples beat adjectives. "Professional but warm" means nothing. Three real emails or blog posts mean everything.
Rules. "One em dash per document. No 'it's not just X, it's Y.' No bullets unless I ask."
Backbone. "Take a position. Cut hedges that protect the writer. Keep caveats that protect the reader." The second half matters at a financial services company. Some caveats are load-bearing.
Reality. "Ask me for the specifics you're missing instead of inventing them."
Pushback. "If my premise is wrong, say so before you write a word."
The prompt I use.
The prompt I use most isn't a drafting prompt at all:
"Edit this page for signal-to-noise: cut throat-clearing, hedges, and restated context, but preserve every claim, number, name, link, and decision. If unsure whether something is load-bearing, keep it. Return the edited page plus a note on anything you cut that might matter."
Compare that to "make it shorter," which hands the model the decision about which facts survive. This one names what's protected, defaults to keeping anything ambiguous, and reports what it cut so I can audit the call. I think of it as compression with a receipt.
Your human edit is the whole job.
The draft is the machine's but the judgment is always yours. Here’s my checklist:
Read it out loud. If you wouldn't say the sentence, cut the sentence.
Cut restated context, wherever it lives. If your opening paragraph is the answer, keep it. If it's a warm-up, cut it.
Check every specific. Names, numbers, dates, links. All of them.
Add what only you know. The client's name, the deal that didn't close, or the real story because specifics are the human fingerprint.
Kill the anywhere-sentences. If a line could appear in anyone else's memo, it isn't earning its place (just like a company value!).
You're the byline. If you didn't edit it, don't send it.
Try it on something you published.
We built an exercise for a lunch and learn at Lineage. Take a page you actually published, ask a model to redraft it "polished and professional," and hunt. Ours came back with template headers ("Unlocking the Power of..."), a journey or two, a hedge stack, and one change almost nobody spotted at first: the transaction fee quietly moved from $749 to $479 and even I nearly missed it. AI tells in the copy and a number that mattered. Damn.
AI drafts. You decide. Ship nothing you haven't made yours.
AI: 60% | Human: 40% — Claude drafted the post from the lunch-and-learn deck it and Jesse built together. Jesse set the topic and taxonomy, red-teamed the source deck, wrote the opening, and contributed the edit-pass prompt.
