How to Make Claude Write Like You: Build a Banned-Phrases Voice Spec

TL;DR: Paste in samples and ask Claude to “sound conversational” and you get the median internet voice. Keep a running list of what you would never say, load it before every draft and check against it after, and you get a lot closer to yours. The list only grows if you read your drafts and write the failures down.

“Navigate the complexities.” That’s the phrase that stopped me. Three drafts deep on a piece about apartment fermentation, one in the morning, Linux Mint terminal open, and there it was. The sentence looked fine. I almost let it through.

It wasn’t mine. My grandfather ran a sawmill in Tennessee and he never said it. Nobody at the truck-stop counter where I read paperback Bukowski over thirty-cent coffee said it. Nobody in the smoke pit on the flight line at Bagram said it either. I’ve never stood in a room where that phrase lived. Claude had it ready anyway.

Why does Claude’s writing sound like everyone else’s?

Because when you name a style, Claude gives you the most common version of it. “Conversational, personal, no guru energy” points at the middle of every piece of text that ever wore those labels. You’ve read that middle. It’s the LinkedIn coach. It’s the creator-podcast cadence. It’s the paragraph you could unscrew from one blog, bolt onto another, and nobody would notice.

The usual fix is:

• Paste in your past work.
• Add a three-line voice profile to your instructions.
• Ask for “authentic” and “human.”

All of that shows the model what you sound like and asks it to fill in the blanks. Filling in the blanks means averaging. Averaging is what made the median voice in the first place. Feed it more samples and it just averages harder.

What fixes it?

A list of the words, sentence shapes and moves that don’t belong in your mouth. “Be authentic” gives the model nothing to do. “Never write ‘navigate the complexities’” is a rule it can actually follow. Enough of those rules fence off everything that isn’t you.

What goes on the list:

• Phrases you caught in a draft and flinched at.
• Sentence shapes that read as AI the second you see them, like the mirrored “this isn’t X, it’s Y.”
• Office language you picked up and never washed off. Mine came from six years at GoDaddy and two more at a real estate firm where everybody talked about stakeholders.
• Motivational tag-lines and neat little closers.
• Openings you’d never use.

Nobody can hand you this list. Mine is tuned to a sawmill grandfather, a flight line and a Cuban-American household. You came up somewhere else, so your list will look different.

What’s actually in a voice spec?

Mine runs around six thousand words across twenty sections. Most of it is grievances. The parts that pull their weight:

• Banned vocabulary. Added one phrase at a time over a year of catching tells.
• Banned structures. The “this is not X, this is Y” pattern has its own catalog, because Claude tries to sneak a new variant past me almost every week.
• Openings I never use.
• A profanity budget. It changes by piece type and by section.
• Seven literary touchstones. Each one is named along with the specific structural move I want from it. A name alone doesn’t tell the model what to do.
• Past failure points. The exact ways the voice has fallen apart before, so the model can scan for them.

How do you wire it into Claude?

Put the short, always-true rules in your account preferences and the full spec in a skill. Preferences apply to every conversation you have. Project instructions apply only inside that project, which is handy if the spec belongs to one book or one client. A skill loads when Claude decides your request needs it, so a six-thousand-word spec doesn’t ride along when you’re asking about a recipe.

Two things to check before you blame the skill:

• Code execution has to be on. Skills need it. On a personal plan it’s under Settings, Capabilities, and the skill toggles live under Customize, Skills. On Team and Enterprise an owner turns it on for the org first.
• The description does the picking. Claude reads a skill’s description to decide when to use it. Write that field like a trigger (“use before drafting any post, essay or newsletter”), not a tagline. If it still won’t fire, name the skill in your prompt.

I run the spec through two skills.

1. Pre-flight runs before anything gets drafted. It loads the spec and makes the structural calls for the piece: what kind of piece it is, the emotional register, where the profanity goes if any goes anywhere, and the shape of the ending so the prose has somewhere to head. It also sets the metaphor lane. I pull from biology, chaos magic and military logistics, and keeping it there keeps out the worn-out creator-economy imagery the model grabs by default.
2. Post-flight runs on the finished draft. Six passes: kill the AI tics, audit the voice against the spec line by line, cut generic language, catch repetition, check rhythm, and run a perplexity pass to break up the flat cadence the model slides into when it’s coasting. Then one last read of the whole thing as a piece of writing instead of a checklist.

Pre-flight tells the model what I don’t say. Post-flight catches what it said anyway. Whatever gets past both goes into the spec.

Does it really change the output?

Yes. Same model, same prompt. The only difference is what got loaded before the draft. Here’s stock Claude on whether to niche down, paraphrased:

In the creator economy, many people grapple with whether to niche down. Experts advocate focusing on a single lane, but there’s a compelling counter-argument: diverse interests can build something more authentic and sustainable.

And here’s the same prompt run through the spec:

Niche-down advice is the safest thing a coach can sell you. Pick one lane and build authority and the expertise will compound, supposedly. The trouble is most of us are not one thing.

The first one could have come from anybody. The second one takes a side, and you can argue with it. If you can’t argue with it, it isn’t saying anything.

How do you keep the list growing?

Every time you read a draft, look for the one line you would never say and add it to the list. I add to mine almost every week. The model will crank out ten thousand drafts without complaining, and it will never notice when one of them is dead. That part’s on you. Whatever you notice has to get written down somewhere the model will read it next time.

When I caught “navigate the complexities,” I added it and the spec went from 5,847 words to 5,851. The next morning the model didn’t reach for it.

Frequently asked questions

Can I just give Claude my writing samples instead?

Samples help, but on their own they get blended with everything that looks like them, and you get a smoothed-out average of yourself. Use them for rhythm and pair them with a list of what to avoid.

Should the voice spec go in preferences, a Project or a skill?

Put rules that hold everywhere in your account preferences. Put the full spec in a skill, which loads only when Claude decides the task needs it. Use a Project when the spec belongs to one body of work, like a single book or client.

How long should a banned-phrases list be?

Start with ten lines. Mine is around six thousand words across twenty sections, but it got there one catch at a time over a year. Don’t try to write it in one sitting. The length comes from paying attention.

Will this work with models other than Claude?

The idea works with any model, since they all drift toward the average of what they were trained on. The setup is different. In Claude it’s preferences, Projects and Skills, and other tools have their own versions.

Will a spec stop Claude from sounding like AI entirely?

No. It closes the gap, and catching new tells every week keeps closing it. The model keeps inventing new tells too, which is why my list of banned structures keeps getting longer.

Can I use someone else’s voice spec?

Borrow the categories. The entries have to come from you, because what doesn’t belong in your mouth depends on where you’ve been.

---

If you want the same grow-one-file-over-months approach applied to a whole catalog, that’s Thirty Bricks.