Claude can write. That was never the problem. The problem is that its default voice is recognizable in about two sentences — hedged, evenly paced, allergic to a real claim — and once a reader clocks it, they stop trusting the content. This is the fix I use on everything that goes out under my name: the tells to cut, the words to ban, the swaps to run, and how to get Claude writing the way you already write.
Stop treating this as a vocabulary problem. The words are downstream. Claude hedges because it was trained to be safe. It writes three-item lists because three-item lists score well on average. It uses transitions because it can't trust you to follow without them. Fix the structure and most of the vocabulary fixes itself.
Every tell lands in one of eight categories. Learn the categories, not the word lists. The specific words shift every few months. The categories haven't moved in three years.
The burned words. If these survive into a final draft, nobody edited it: delve, tapestry, testament, realm, landscape, paradigm, robust, leverage, utilize, facilitate, harness, unlock, elevate, foster, embark, underscore, multifaceted, holistic, seamless, paramount, meticulous, pivotal, crucial.
The quiet wave. Less obvious, equally fatal, and they leak across a whole draft: "that said," "at its core," "when it comes to," "in today's fast-paced world," "game-changer," "in the realm of," "by doing X, you can Y," "whether you're A or B," "it's worth noting."
The compensation wave. This is the 2026 one, and it fools people who know the first two lists cold. The models learned that the burned words get caught, so they overcorrect into manufactured casualness. "Honestly," dropped in for false intimacy. "Look," as an opener. "Here's the thing:" as a transition. "Hot take:" in front of an opinion the model doesn't hold. Fake I-learned-this-the-hard-way stories with no specifics in them.
None of these words are bad on their own. They're banned because they've been used so heavily that they now read as a handshake between one model and another. Cut them unless you have a specific reason to keep one.
However. Therefore. Thus. Hence. Indeed. In essence. In summary. In short. Plus significant, substantial, and notable any time there's no number attached to them.
Banning a word is only half the job. You need somewhere to go. These are the substitutions I run first, before touching structure, because they're mechanical and they take about ninety seconds.
| Instead of | Write |
|---|---|
| delve | examine, dig into, look at |
| leverage (as a verb) | use, build on, apply |
| utilize | use |
| facilitate | run, lead, help, set up |
| harness | use, apply |
| unlock | access, open, get to |
| elevate | raise, improve, upgrade |
| foster | build, grow, encourage |
| embark | start, begin |
| underscore | stress, repeat, emphasize |
| paramount | most important, top priority |
| meticulous | careful, precise, exact |
| robust | strong, solid, holds up |
| seamless | smooth, no friction |
| multifaceted | name the facets |
| holistic | whole-system, end-to-end |
| pivotal | key, decisive |
| crucial, critical | must, required, key |
| arguably, notably, interestingly, ultimately | cut entirely |
| furthermore, additionally | and, also |
| moreover | cut entirely |
| Instead of | Write |
|---|---|
| in today's fast-paced world | cut entirely |
| it's important to note that | cut it and just say the thing |
| it's worth mentioning that | cut entirely |
| that said | but, still |
| at its core | fundamentally, the basic issue |
| when it comes to | for, on, in |
| in the realm of | in |
| a tapestry of | a mix of, a set of |
| a testament to | proof of, shows |
| navigate the landscape | deal with, work through |
| paradigm shift | rethink, change in approach |
| game-changer | cut it and describe what changes |
| a plethora of | a lot of, or an actual number |
| by doing X, you can Y | X gets you Y |
| whether you're A or B | for both A and B |
| dive deep into | look at, study |
| uncover insights | find out, discover |
| unleash the power of | use |
| at the end of the day | in the end, or cut |
| it goes without saying | cut entirely |
| in conclusion | cut it and let the ending land |
| AI version | Human version |
|---|---|
| leverage our strengths | use what we're good at |
| delve into the data | look at the numbers |
| seamless integration | fits in cleanly |
| meticulous attention to detail | careful work |
| harness the power of AI | use AI |
| foster collaboration | get people working together |
| by leveraging automation, you can save time | automation saves time |
| navigate the competitive landscape | compete |
| a holistic view | the full picture |
| interestingly, the data revealed | the data shows |
| in conclusion, we recommend | we recommend |
Everything in the two previous sections is subtraction. Subtraction alone produces clean, competent, forgettable prose. These six laws are the addition side, and they're where the actual voice comes from.
Numbers, names, dates, places, dollar amounts, exact quotes. "Many companies struggle with hiring" is dead on the page. "We posted the PM role, interviewed 43 people, three of the top five turned us down over comp, filled it in six weeks from our own network" is alive. Specificity is the cheapest and fastest signal that a person wrote this.
AI writes 15-to-18-word sentences at a metronome pace. People cluster: four words, eight words, thirty-two words, six words. Vary the length, the opening, the structure. If you can clap a steady beat through three paragraphs, a machine wrote them.
AI hedges because it was trained not to be wrong. People rank things, judge things, name names, and put their reputation behind it. Every paragraph should carry at least one claim someone could argue with.
Fragments. Sentences that open with a conjunction. One-word paragraphs. Run-ons that build pressure. AI breaks rules by accident and inconsistently. People break them with intent, and the break carries the meaning.
Real thinking is messy. Writers reverse themselves mid-paragraph, drop a clause, leave a contradiction standing, end on a question. Prose that's perfectly smooth and fully resolved reads as machine-finished. Leave some seams visible.
AI writes about every subject the same way. You don't. You return to certain ideas, certain framings, certain words. Don't sand that down. The repetitions are the voice.
Most people try to explain their voice to Claude in a sentence or two. "Write it casual but professional." That produces nothing, because you're describing a feeling and Claude needs patterns — how long your sentences run, how you open, how you sign off, the words you reach for and the ones you never touch.
You already have thousands of samples of your own writing. Point Claude at them.
Connect Gmail (or Outlook), then run this once:
Thirty is roughly the floor. Below that Claude picks up noise instead of pattern. Sent mail beats anything else you could feed it, because it's writing you actually shipped to a real person under a real deadline, rather than writing you polished for an audience.
The profile comes back as a list of observations, and this is the part people skip. Read it. Some of what it finds will be habits you don't want reinforced — I had a stretch of over-apologizing in email that Claude dutifully wrote down as a signature move. Delete those lines. What's left is your voice, minus the tics.
A voice profile sitting in a chat window is worth nothing next week. Put it in the file Claude loads at the start of every session — the next section covers exactly where that lives and what else belongs in it.
Everything above is worthless if you have to remember it each time. The point of a rules file is that Claude reads it before it writes a word, every session, without you asking.
CLAUDE.md file at the root of your project folder. Loaded automatically at the start of every session in that folder.Copy this whole thing. Swap in your own voice profile from Step 5 where marked.
For a session where you don't want to set anything up, this does most of the work on its own:
Run these on anything going out under your name. Most drafts fail two or three, and fixing those two or three is usually the whole job.
| Test | How to run it |
|---|---|
| Em-dash count | Per 500 words. AI averages 4–9, people average 1–3. Over 5 means half need to become commas or periods. |
| Hedge count | Count "it's important to note," "however," "arguably," "could," "may," "in many cases" per 500 words. Three or more is the AI hedge pattern. |
| Subject-first ratio | How many sentences open with "The," "This," "It," or "Our"? AI runs 60%+. People run 30–40%. Reorder until you're under 50%. |
| Paragraph variance | All paragraphs three or four sentences? That's the AI default. People cluster 1, 2, 5, 1, 7. |
| Connector density | "furthermore," "moreover," "additionally," "in conclusion." One or more per paragraph reads as AI. |
| Adverb density | Words ending in -ly. AI runs 2–3 per paragraph. Target 0–1. Cut "significantly," "ultimately," "essentially," "basically" first. |
| Passive ratio | "is / was / been / are" plus a past participle. AI runs 20–30%. Target under 10%. |
| Three-item count | Count the "A, B, and C" lists. Convert at least one to a pair, or split it into separate sentences. |
| Vague-noun audit | "area," "space," "landscape," "realm," "domain," "opportunity" with nothing specific attached. Each one needs a name, a number, or an example. |
| Claim-to-evidence ratio | Every claim should have evidence next to it — a number, a name, an anecdote. AI drifts to three claims per piece of evidence. |
| The Slack test | Paste a paragraph into Slack to a colleague. Would they reply "...what?" If so, it's too formal. |
| The reorder test | Shuffle three paragraphs. If the new order reads as well as the old one, the prose has no rhythm. |
If you read nothing else here, read this.
Don't try to fix a draft in one pass. Cut before you add, structure before you polish.
| Pass | What you're hunting |
|---|---|
| 1. Hedge cut | Every weakening qualifier. If the sentence becomes a confident claim without it, the cut was right. |
| 2. Banned-word swap | Run the swap list. If there's no clean replacement, cut the sentence. |
| 3. Connector cull | Delete 60% of the "howevers" and "furthermores." Replace with a period. |
| 4. Opening variation | Read the first three words of every sentence. Three in a row starting the same way? Rewrite one. |
| 5. Length burst | Add at least two sentences under 8 words and one over 28. |
| 6. Specificity injection | Every vague noun and metric-free claim gets a name, number, or date. |
| 7. Claim density | Every paragraph needs one arguable claim. Pure-background paragraphs get folded in or killed. |
| 8. Voice injection | One human signal per 200 words. A fragment, a contradiction, a specific story. |
| 9. Read aloud | Rewrite everything that trips. |