How to make Claude write like a human: styles, skills and Claude Code

To make Claude write more like a human, focus on the characteristics of Claude that make it seem less human, not the ones you'd fix in ChatGPT. Claude hardly uses em dashes, but it overuses phrases like "plenty of", "roughly", "here's" and "rather than". When it edits, it rarely begins a sentence with And, But or So, and its sentences tend to be the same length. Put these as rules in Instructions for Claude, a project, a skill or a CLAUDE.md file. Or use a different model to humanize Claude's work over MCP.

Does Claude sound like AI in the same way ChatGPT does?

No. It sounds like AI in its own way. So, if you're trying to prompt Claude to write better, prompts that work on ChatGPT don't work that well here. The numbers below come from two studies: one where ten models from 2026 polished 7,554 passages of human writing (2,463 of those rewrite pairs used Claude Opus 5.5) and another where four models from 2026 wrote 1,177 pieces of writing from scratch (Anthropic's model for that was Claude Sonnet 5.5). These are different models, so don't compare their numbers against each other.

Punctuation. Claude doesn't use many em dashes, so you don't need to tell it not to. When rewriting human writing, Opus 5.5 used 0.10 per 1,000 words, compared to 2.56 for the human bloggers whose writing it was rewriting, and 3.88 for OpenAI's GPT models. When writing from scratch, Sonnet 5.5 used 0.09 per 1,000 words, the lowest of the four models tested, and much lower than the 0.72 in the human originals.

Phrases. "Plenty of" is very common (0.68 per 1,000 words vs 0.07 for the humans). Claude also leans on "roughly", "genuinely" and "isn't". Across all the models, "genuinely" and "roughly" were used 12.5 and 7.7 times as often as by the humans, and Claude is one of the reasons why. Writing from scratch, Sonnet 5.5 had its own favorites: "here's" (0.45 per 1,000 words) and "rather than" (0.39), against 0.08 and 0.07 for people.

Starting words. It's also uncommon for Claude to begin sentences with "And", "But", or "So". When rewriting human writing it did so 5.0% of the time, the lowest of any model in our dataset (OpenAI's models did so 9.7% of the time; the human bloggers did so 12.0% of the time). When writing from scratch, where the human documents span 20 kinds of writing, humans began sentences this way only 3.8% of the time, while the AI models did so 3.4% of the time, so the bloggers' 12.0% was partly their own habit.

Sentence length. When rewriting, Claude's sentence lengths are more uniform than OpenAI's (0.56 vs 0.59 in coefficient of variation; the bloggers were 0.65). This gives the writing a careful and tidy feel, which is the problem. People aren't that tidy. When writing from scratch, the gap was smaller, and the difference between Claude Sonnet 5.5 (0.49) and humans (0.50) was less than for GPT-5.4 mini (0.43). Note that the sentence length numbers aren't comparable between writing from scratch and rewriting human writing, since they're using different sets of texts and measuring things slightly differently.

Formatting. When writing from scratch, all four AI models added a lot of unasked-for formatting to their writing. On average they used 10.1 bolded phrases, 4.2 headers and 8.9 lines of bulleted lists per 1,000 words, compared to 2.4 bulleted lines in the human originals. The formatting showed up even in emails.

What instructions should I give Claude?

Here are some rules based on what Claude tends to do. You can use this block with any of the options in the next sections.

When you write something other people will read as mine (emails, posts, announcements):
- Avoid "plenty of", "roughly", "genuinely", "truly", "essentially", "rather than" and "for instance". Use a number, "about", or just the example.
- Don't introduce things with "Here's". Just say them.
- Vary sentence length a lot. Put a short sentence after a long one. Fragments are fine.
- Start some sentences with And, But or So, the way people do.
- Watch sentences built on "isn't". If you're correcting a claim nobody made, cut it and state the point.
- Don't hedge every claim. If I gave you the fact, state it.
- No headers, bold or bullet points unless I ask for them. Write prose.
- Match the samples of my writing in this project or conversation over your own defaults.

We don't have a rule for em dashes because they're not a problem for Claude. The "Here's" and "rather than" lines are from Claude Sonnet 5.5 drafts, and the formatting line is there because formatting was the biggest tell when models were writing from scratch. The lines about starting sentences with conjunctions and varying sentence length are the most important, because they go after the rhythm that word bans can't touch. The "isn't" line is our own advice, not a measured finding: we think a lot of the time a sentence with "isn't" sets up a contrast nobody asked for ("This isn't about speed. It's about trust."). See our guide on "it's not X, it's Y" for more on that pattern.

Can I use Claude Styles to make it sound like me?

That may no longer be possible. Styles were the original way to customize Claude's writing style, either using one of the presets or uploading a couple of examples of your own writing to create a custom Style. As of November 2026, Anthropic's help article on Styles no longer loads, and their article on how to personalize Claude only shows three features: Instructions for Claude, project instructions and skills. If you still have Styles, you can add this block there, and it will have the same effect as adding it to Instructions for Claude. If not, use one of these.

Instructions for Claude change Claude's writing in every conversation. Click your initials in the bottom left, then Settings, then Your Claude (or Account in the older layout), and paste the block above under Instructions. Keep it short, since it applies to everything, including code questions and quick lookups.

Project instructions only apply to one project, so they're the better home for writing rules. If you create a project for the writing you want to sound more human (whether that's client emails or LinkedIn posts), you can add this block to the project instructions along with three or four samples of your own writing, so Claude has something to copy.

Should I use a humanizer skill?

It's a good free option, and it might be all you need. A skill is a folder containing a skill.md file with instructions that Claude loads when the task calls for it. In Claude.ai, you'll find skills under Customize, then Skills (code execution has to be turned on to use them).

An alternative, if you don't want to write your own, is to use blader/humanizer, an open-source (MIT license) skill from a third-party developer. It's based on the Wikipedia guide Signs of AI writing and includes 26 different patterns (from the "not X but Y" phrasing, to one-liner closers, to run-ups where the text talks around something before getting to what it wanted to say), and tells the agent to rewrite them out. It's very popular, it's free, and it's really good. To use it in Claude.ai or the Claude desktop app, download the repository as a ZIP file and upload it as a skill in Settings. For Claude Code, see below.

The reason for the limit is that skills are, fundamentally, Claude editing itself. It will catch the patterns it's told to look for, but it's going to edit in very similar ways to how it wrote in the first place, which brings back that even rhythm and safe word choice. If you'd like to learn more, check out why your AI can't fix its own writing.

How do I make Claude Code write like a human?

With a CLAUDE.md rule, a skill, or both. Claude Code reads CLAUDE.md files at the start of every session. Use ~/.claude/CLAUDE.md in your home directory if you want a rule to apply to every project, or ./CLAUDE.md in a project to share it with your team. Since you probably want Claude Code to write like a human whenever it writes (READMEs, release notes, customer emails), add the block above to your user file.

Anthropic's docs make a point worth knowing: Claude treats CLAUDE.md as context, not enforced configuration, and the more specific and concise your instructions, the more consistently it follows them. A short list of named words beats a paragraph about tone.

To add blader/humanizer to Claude Code, run the commands from its README:

/plugin marketplace add blader/humanizer
/plugin install humanizer@humanizer

Then run the skill in any session with /humanizer:humanizer. This needs Claude Code 2.1.142 or newer. On an older version, use npx skills add blader/humanizer --global --agent claude-code instead.

What does an MCP humanizer do that a skill can't?

It uses a different model to edit your writing, instead of Claude editing itself. When you give Claude a prompt, style, project or skill, that's additional instructions for how Claude should write something, but Claude still does the work. When you use an MCP server tool like Zavus, it's sending the text Claude generated to a completely different model, built only for rewriting, that is then rewriting it and sending it back to Claude for Claude to check and send to you.

Because the model rewriting your writing doesn't have the same preferences as Claude, it won't add in "plenty of" or even out sentence lengths. Below is an example of a paragraph Claude generated for a request to polish a piece of writing, and then the same paragraph polished by Zavus.

To add Zavus to Claude Code, run:

claude mcp add --transport http zavus https://api.zavus.ai/mcp

Then run /mcp in your Claude Code session, select zavus and sign in. In Claude.ai or Claude Desktop, add it as a custom connector under Customize, then Connectors. Steps for every other client are in how to add a humanizer to Claude, ChatGPT, Cursor and Claude Code.

Which option should I pick?

Start with the free options. If you're quickly writing something for yourself, adding the block above to Instructions for Claude should be enough. For writing you do every week, set up a project with your own samples. In Claude Code, add both a CLAUDE.md rule and the blader/humanizer skill.

If the writing goes out under your name (to clients, customers or investors) and it still sounds too much like Claude after all of that, try a separate model. Our guide to making AI writing sound human explains the trade-offs, and Zavus for founders shows the kind of writing it's built for.