Claude humanizer: 8 tells to fix and how to rewrite them

1. TELL 01

   Em dashes as glue

   A July 2026 Economist study, cited on Wikipedia, found only Claude used em dashes more than professional writers. Graphite later put Claude Opus 5 at about the human rate, so older drafts show it most.

   Example we wrote

   Claude-style

   The draft was ready — or so we thought — until the client called.

   Rewritten

   The draft was ready, or so we thought. Then the client called.

2. TELL 02

   'Not X, it's Y'

   Wikipedia lists negative parallelisms as a classic sign. When PCWorld had Claude audit a writer's habits, "it's not just this; it's that" made the list.

   Example we wrote

   Claude-style

   This isn't just a budgeting app. It's a new way to think about money.

   Rewritten

   The app shows where your money went last month, sorted by how much you regret it.

3. TELL 03

   'Less X, more Y'

   Graphite measured "less X, more Y" comparisons in Claude Opus 5 at 82 times the human rate, and "rather than merely" at 160 times.

   Example we wrote

   Claude-style

   Think of onboarding less like a checklist and more like a conversation.

   Rewritten

   Skip the checklist on day one. Ask the new hire what they want to learn first.

4. TELL 04

   Stacked superlatives

   Graphite found 'single most' and 'arguably the most' in Claude but never in GPT-6 Astra, and 'every single' at 112 times Astra's rate.

   Example we wrote

   Claude-style

   This is arguably the single most important step in the entire process.

   Rewritten

   Do this step first. The next three depend on it.

5. TELL 05

   Hedged openers

   "It's worth noting" openers were another habit Claude flagged in PCWorld's test. Hedges that point at no real doubt read as filler.

   Example we wrote

   Claude-style

   It's worth noting that results may vary depending on a number of factors.

   Rewritten

   Large files take longer, and slow Wi-Fi makes it worse.

6. TELL 06

   Inflated vocabulary

   Delve, pivotal, underscore, tapestry and testament all sit on Wikipedia's words-to-watch list, along with "serves as" and "boasts" where "is" and "has" would do.

   Example we wrote

   Claude-style

   The launch underscores the team's pivotal role in a rich tapestry of innovation.

   Rewritten

   The launch shows the team can ship on time.

7. TELL 07

   Chat leftovers

   A Claude Code GitHub issue titled "Claude says 'You're absolutely right!' about everything" drew 870 thumbs-up. Praise and sign-offs from the chat don't belong in a final draft.

   Example we wrote

   Claude-style

   Great question! Here's a clear breakdown of the key points. I hope this helps!

   Rewritten

   Three things changed in the new contract.

8. TELL 08

   Summary closers

   Wikipedia flags formulaic endings such as 'Despite these challenges' wrap-ups. Claude often ends by restating what the reader just finished reading.

   Example we wrote

   Claude-style

   Ultimately, by following these steps, you'll be well on your way to success.

   Rewritten

   Send the invoice the same day you finish the job.

1. 01

   Read it out loud once

   Your ear catches the sing-song rhythm faster than your eye. Mark every line you'd never say to a friend.
2. 02

   Cut the opener and the closer

   Delete warm-up lines and the final paragraph that restates everything. Start at the first useful sentence and stop after the last one.
3. 03

   Fix the em dashes case by case

   Keep one if it earns its place. Turn the rest into commas, periods or parentheses, whichever matches the pause you want.
4. 04

   Flatten 'not X, it's Y' lines

   If nobody claimed X, just state Y. Corrective framing makes the reader feel argued with for no reason.
5. 05

   Swap the inflated words

   Delve, pivotal, underscore, tapestry, testament and showcase are on Wikipedia's watch list. Use the plain word you'd say out loud.
6. 06

   Break up lists of three

   Claude likes three adjectives or three short phrases in a row. Keep the one that matters, or give a real example instead.
7. 07

   Delete empty hedges

   'It's worth noting' and 'results may vary' rarely add anything. Keep a hedge only when you'd name the real reason for doubt.
8. 08

   Add one thing only you know

   A name, a number from your own work, a mistake you made. Specifics are what generic drafts can't supply.
9. 09

   Vary the sentence length

   Mix a short line with a longer one. Uniform medium sentences are part of what makes AI prose feel flat.
10. 10

    Strip the chat formatting

    Bold lead-ins, headers on a short answer and title-case bullets come from chat replies. Turn them back into sentences for emails and posts.

Write a short blog post about [your topic] for [your reader].

Write in smoothly flowing prose paragraphs, the way you'd explain this to a colleague.

Here are a few paragraphs I wrote. Copy their rhythm and word choice:
[paste your sample here]

Use commas or periods where you'd use an em dash. If nobody claimed the opposite, state the point directly instead of saying what it isn't.

Read the full Claude prompting best practices for format control, and PCWorld's write-up for an example of Claude ranking a writer's own AI-sounding habits.

Also cited: the Claude Code sycophancy issue and The Register's coverage of it.