You wrote something with Claude. You read it back. And somewhere around the third em dash and the phrase "let's dive in," your stomach dropped, because it sounded exactly like the robot slop you were trying to avoid making in the first place.
I feel this in my bones. I once spent a whole afternoon getting an automation to sound less like a brochure that had somehow learned to talk, which is a sentence that says a lot about how I spend my Thursdays. But here is the good news: Claude sounds like AI because nobody told it not to. It is a very eager golden retriever. It will fetch whatever you throw, and by default it throws back "corporate blog post energy" because that is what most of the internet trained it on.
You can fix that in about five minutes. Let me show you the tells first, because you cannot un-see them once you know them, and then I will hand you a block of text to paste in that turns most of it off.
The Tells That Give It Away
Here is what "wrote by AI" actually looks like on the page. Once you spot these, every LinkedIn post you scroll past will suddenly light up like a bad Christmas sweater.
The em dash. The little horizontal stick that Claude sprinkles everywhere like it is being paid per dash. Real people use commas. Or parentheses. Or they just write two sentences like a normal human who owns a keyboard.
The "it's not just X, it's Y" move. "It's not just a coffee maker, it's a morning ritual." AI loves this pattern the way I love pizza, which is to say relentlessly and past the point of good judgment. It shows up constantly and it always sounds like an ad for a mattress.
The vocabulary that fools nobody. Delve. Leverage. Utilize. Optimize. Streamline. Robust. Scalable. Seamless. Tapestry. Testament. Nobody talks like this at dinner. If your neighbor said "let's leverage the grill to optimize our robust cookout," you would quietly move.
The throat-clearing. "It's important to note that." "It's worth noting." "That said." These are the verbal equivalent of clearing your throat for ten seconds before saying anything. Just say the thing.
The suck-up opener. "Great question!" You did not ask a great question. You asked a normal question. The compliment is filler and everybody can smell it.
The wrap-up nobody needs. "In conclusion." "In summary." You are not writing a book report in 1994. When you are done, stop. That is the ending. You just do it.
The rule of three, forever. AI has a thing for listing exactly three items, every time, in every sentence, because triads sound smooth and complete and balanced. See what I did there. It gets predictable fast.
The relentless cheerfulness. Every sentence beaming at you like it just got a gold star. Warmth is good. A guide that reads like a hostage being forced to enjoy a birthday party is not warmth, it is a cry for help.
The Paste-In Fix
Here is the actual tool. Drop this into your custom instructions, the settings field where you tell Claude how to behave across all your chats. Then Claude carries these rules into everything without you nagging it every single time.
Write like a real person, not an AI assistant. Specifically:
- No em dashes, ever. Use commas, parentheses, or two separate sentences.
- Never use the "it's not just X, it's Y" pattern or any version of it.
- Banned words: delve, leverage, utilize, optimize, streamline,
robust, scalable, seamless, tapestry, testament.
- No throat-clearing. Skip "it's important to note," "it's worth
noting," and "that said."
- No suck-up openers like "Great question." Just answer.
- Don't end with "In conclusion" or "In summary." Stop when you're done.
- Vary sentence length. Mix short punchy lines with longer ones.
Some sentences should be five words.
- Take a position. Don't hedge every sentence into oatmeal.
- Dial the enthusiasm way down. Warm is fine. Cheerleader is not.I do not have the exact menu path memorized, and Claude's settings move around often enough that I would be guessing, so go poke around in Settings and look for the field about personal preferences or how Claude should respond. Confirm the current spot in the app before you trust it. (If you want the deeper version of this, a saved writing Style trained on your own samples does an even better job, but that is its own guide and its own rabbit hole.)
Your Nuke Words
The banned list up in that block is the generic stuff every AI overuses. But you almost certainly have your own words that make your skin crawl, the ones that instantly sound like a press release wearing a person costume. I call mine nuke words. They never survive contact with anything I publish. Not once. Not "just this time."
Here is my starter set, the words that get nuked on sight. Steal the ones that make you wince, skip the ones you actually use, and add your own until the list feels like yours.
delve, leverage, utilize, optimize, streamline, robust, scalable, seamless, tapestry, testament, elevate, unlock, unleash, unveil, embark, foster, harness, realm, myriad, plethora, beacon, captivate, resonate, bustling, vibrant, holistic, nuanced, meticulous, pivotal, profound, compelling, transformative, groundbreaking, cutting-edge, ever-evolving, game-changer, supercharge, empower, delight, demystify, navigate, landscapeThen there are the phrase tells, the little stock openers and filler that AI reaches for on autopilot. Kill these too:
it's important to note, it's worth noting, in conclusion, in summary, in the realm of, when it comes to, gone are the days, look no further, the ultimate guide, we've got you covered, a testament to, paving the way, in today's digital age, not just X but alsoPaste your version right under the block from the last section so Claude carries both at once. One will still slip through now and then, because these words are sneaky little raccoons that get into everything. That is exactly what the next check is for.
Give Claude a Voice to Aim At
Getting rid of the tells stops Claude sounding like a robot. It does not make Claude sound like you. Those are two different jobs, and the second one needs a target. So hand Claude a short profile of the voice you want: the personality, the kinds of jokes, the way you actually talk.
Here is mine. This is my real voice profile, the one I feed Claude for this whole series, so what you are reading right now came out the other end of it. Yours should look nothing like this, and that is the entire point. Treat it as an example of the kind of thing you can add, not a template to copy word for word (unless you also want to be a squirrel, in which case, pull up a chair).
The Caffeinated Squirrel. I explain automation like someone who has had exactly the right amount of coffee and is genuinely excited about this stuff, while also being fully aware that caring this much about Make.com scenarios is objectively ridiculous.
In bounds, use freely: groan-worthy puns (the worse, the better), self-deprecating asides about being automation-obsessed, absurd comparisons that somehow make the technical thing click, food and pizza analogies, the occasional 80s or 90s wink, and dad jokes, because I have three kids and it is now a personality trait.
Out of bounds, never: death, sex, dark humor, current politics, anything mean toward a specific person or company, and the kind of cynicism that makes the actual thing sound worthless.
Warmth without cheerleading. I would rather make you laugh than make you feel motivated.
Drop your version into the same settings field, or into a specific Project if that voice only belongs to one kind of work. Now Claude has something to aim at instead of defaulting to helpful-robot beige.
The Read-Aloud Check
Here is the one that catches everything the paste-in block misses. Read the output out loud. Actually out loud, with your actual mouth, ideally when nobody is watching because you will feel ridiculous and that is fine.
If you hit a sentence you would be embarrassed to say to a friend over coffee, that sentence is still AI. Cut it or rewrite it the way you would actually say it. Your ear catches what your eyes skim right past. The word "robust" reads fine on screen and then dies in your mouth the second you try to say it out loud. Your ear is the better editor. Let it win.
Where This Goes Sideways
A few honest warnings, because this is not a magic force field and I would rather you know now.
It drifts in long chats. Forty messages deep, Claude starts backsliding into helpful-robot mode because the conversation buried your instructions under a pile of other stuff. When you catch it, just say "you're sounding like AI again, tighten it up," or start a fresh chat. Fresh chats are the reboot-the-router of AI writing. Weirdly effective.
It only applies where you set it. Custom instructions cover your normal chats, but a Project can run its own rules and override them. If one workspace sounds great and another sounds like a robot, that is usually why. Check the settings on whatever you are actually working in.
You can over-correct into a ransom note. Strip out every ounce of personality and you get prose so clipped it sounds like a threatening letter made of magazine cutouts. The goal was never robotic and cold. It was human. Leave the jokes in. Leave yourself in.
You still have to read it. The block does a lot. It does not do all of it. Pasting rules in and then publishing whatever comes out without reading it is like installing a smoke detector and then deciding you never need to check the kitchen again. Read your stuff. You are the last line.
Do This Today
Open your Claude settings right now, find the personal-preferences field, and paste in the block above. Then grab one thing you already published, feed it back with "rewrite this using my custom instructions," and put the two versions side by side. The gap between them is the exact sound you have been trying to get rid of. Once you see it in your own writing, you will never miss it again.