Stop AI Using Em Dashes — Skills & Prompts

Cleaning text afterwards works, but not producing the tells is better. Install the Claude Skill or paste the custom instructions below and the em dashes, curly quotes and filler openers stop arriving in the first place.

Where: Settings → Capabilities → Skills → Create skill

Download the SKILL.md file below.
In Claude, open Settings → Capabilities → Skills.
Create a new skill and upload the file.
It applies automatically whenever Claude writes prose for you.

---
name: plain-prose
description: Write without the punctuation and phrasing tells of AI-generated text. Use whenever producing prose, documentation, email, articles or any writing intended for a human reader.
---

# Plain prose

Apply these rules to any prose you produce, unless the user explicitly asks otherwise.

Writing style rules:

- Never use em dashes (—) or en dashes (–). Use commas, full stops, parentheses or a colon instead.
- Use straight quotes (" and ') and three full stops (...) rather than curly quotes or the ... character.
- Do not open with "It's important to note that", "In today's world", "In conclusion" or similar filler. Start with the substance.
- Avoid the three-item list as a default rhythm. Vary sentence and clause counts.
- Do not use "delve", "leverage", "robust", "seamless", "landscape", "realm", "testament", "tapestry", "underscore" or "crucial" unless the topic genuinely calls for that word.
- Avoid the "It's not just X, it's Y" and "This isn't about X. It's about Y." constructions.
- Do not end sections with a summarising flourish that repeats what was just said.
- Prefer plain ASCII punctuation throughout so the text survives copy-paste into any editor.
- Vary sentence length deliberately. Some short. Others can run longer when the idea needs the room.
- Write in the register the request implies. Do not default to feature-article prose for a Slack message or a code comment.

## Scope

Applies to prose. Do not apply to:

- Code, where the correct characters are whatever the language requires.
- Direct quotations, which must be reproduced exactly as written.
- Text the user has asked you to preserve verbatim.

## What this does not do

This skill controls writing style only. It has no effect on the machine-readable watermark
Anthropic embeds in generated text, which is applied during sampling inside the model and
cannot be changed by any instruction, skill or system prompt.

A skill is the most reliable option on Claude — it persists across every conversation without re-pasting.

This does not stop watermarking — and nothing prompt-level can

Be clear about what you are buying here. Anthropic's text watermark is applied during sampling, inside the model: "present no matter which Claude product or surface the text comes from", with no switch to turn it off. A skill, style, custom instruction or system prompt is just input tokens. It changes which text gets generated; every token generated still passes through the watermarked sampler. No prompt can prevent it, and any product claiming otherwise is selling you something that does not work.

You cannot use the model to clean the model

The obvious follow-up idea — ask Claude to rewrite its output so the mark is gone — fails for the same reason. The rewrite is itself generated by the watermarked model, so it arrives freshly marked. Only processing that happens outside the model touches the signal: rewriting by hand, translation, or a deterministic tool such as the text watermark remover on this site.

What instructions genuinely do fix

Style, which is the part people actually notice. Em dash density is the single most recognisable signature of machine writing, and it is entirely prompt-controllable. So are curly quotes, the single-character ellipsis, the "It's important to note that" opener, the reflexive rule of three, and the vocabulary tics — delve, leverage, robust, seamless, tapestry. An editor reading your draft reacts to those. No editor has a watermark detector.

Instructions drift, cleaners do not

Model compliance with a style instruction decays over a long conversation, and a skill can be overridden by a more specific request. That is why the sensible setup is both: instructions at the source to catch the bulk, and a deterministic pass afterwards for whatever slipped through. The AI text cleaner is the second half of that, and it never has an off day.

Frequently asked questions

+ Can a Claude Skill remove or prevent the watermark?

No. Watermarking happens during sampling inside the model, so it applies to every token Claude generates regardless of what instructions it was given. Skills, styles, projects, custom instructions and system prompts are all input tokens and none of them can switch it off.

+ How do I stop ChatGPT using em dashes?

Add the rules on this page to Settings → Personalization → Custom instructions, under 'How would you like ChatGPT to respond?'. Compliance decays in long conversations, so if dashes reappear, say 'apply the writing style rules' to reset it.

+ How do I install the Claude Skill?

Download the SKILL.md file above, then in Claude open Settings → Capabilities → Skills, create a new skill and upload it. It then applies automatically whenever Claude writes prose, without pasting anything into each conversation.

+ Is a skill better than cleaning the text afterwards?

For style, yes — not generating a tell beats removing it, and it saves a step every time. But instructions are probabilistic and a model will occasionally ignore them, so a deterministic cleaner as a final pass is still worth running on anything that matters.

More free tools

Claude Text Watermark Remover

Claude's watermark lives in word choice, not hidden characters. Rewrite word choice to disrupt it. Free, runs in your browser.

ChatGPT Watermark Remover

Detect and strip invisible unicode, zero-width spaces and hidden markers from ChatGPT text. Free and entirely in your browser.

Em Dash Remover — Strip Em Dashes From AI Text

Remove em dashes, en dashes and two-em dashes from AI text, or swap them for plain hyphens. Free, instant, no upload.

Invisible Character Detector

Reveal invisible unicode in any text. Lists every hidden code point with its name and count. Free and client-side.

Zero Width Space Remover

Remove zero width spaces, joiners, non-joiners and byte order marks from your text instantly. Free, runs in your browser.

AI Text Cleaner

Strip invisible unicode, em dashes, smart quotes and messy whitespace from AI output in one pass. Client-side, no upload.

Free Text-Cleaning MCP Server

A free hosted MCP server giving your agent text-cleaning tools. Install in Claude Code, Cursor or VS Code in one line.

C2PA Metadata Remover

Remove C2PA provenance metadata, EXIF and XMP from AI-generated images. Re-encodes in your browser; files never uploaded.