A Prompt to Cut Em Dashes and the Rule of Three

Two patterns show up in almost every raw ChatGPT draft, and both are easy for readers to spot before they can say why something feels off. The first is the em dash used as a connector mid-sentence. The second is the habit of listing things in sets of three: "clear, concise, and compelling." Remove those two patterns and the prose reads noticeably more like a person wrote it.

This guide gives you a copy-paste prompt for exactly that job, explains why these patterns cluster in AI output, and shows a before-and-after so you can see the difference.

Why AI Text Uses Em Dashes So Much

The em dash is a versatile punctuation mark. It can replace a colon, set off a parenthetical, signal an abrupt turn, or add drama to a clause. Because it does so many jobs, language models learn to reach for it often. The result is that a single ChatGPT draft might contain eight or ten em dashes in 400 words, which is roughly four to five times the rate you would find in a typical magazine article.

The issue is not that the em dash is wrong. The issue is frequency. When every other sentence uses one to stitch two ideas together, readers sense a stylistic tic without being able to name it. Editors who see a lot of AI output have started calling it the "em dash tell" because it is one of the fastest ways to flag a draft as machine-generated.

To stop ChatGPT em dashes from piling up, the most reliable fix is a targeted rewrite instruction rather than a manual find-and-replace. Find-and-replace swaps the character but leaves the sentence structure unchanged, so the underlying pattern persists. A rewrite instruction tells the model to restructure those clauses entirely.

Why the Rule of Three Follows the Same Pattern

The rule of three is a rhetorical device where ideas are grouped in triplets: "fast, focused, and effective." It works well in speeches and headlines because triplets feel complete and rhythmic. Language models have absorbed enormous amounts of marketing copy and listicles where this structure is everywhere, so they reproduce it constantly by default.

A single use of the rule of three in a 600-word article is fine. Five or six uses starts to feel mechanical, and that feeling is what you want to avoid. Like the em dash, the problem is not the device itself but the density. When you add "remove rule of three prompt" instructions to your workflow, you are asking the model to vary sentence and list lengths so the rhythm feels less rehearsed.

The Copy-Paste Prompt

Paste this as a follow-up message after your initial draft is generated, or include it in a system prompt if you want it to apply to every output in a session.

---

Rewrite the text below. Apply these rules exactly:

1. Remove every em dash. Replace each one with a period, a comma, or a rewritten clause. Do not use a semicolon as a direct substitute unless the sentence genuinely calls for it.

2. Break up any run of three adjectives or nouns listed together with "and." Vary the structure: use two-item lists, single strong adjectives, or rewrite the sentence so no list is needed.

3. Do not add filler transitions ("Additionally," "Furthermore," "It is worth noting that"). Cut them or restructure the sentence.

4. Keep the original meaning, tone, and word count. Do not summarize or add new information.

[PASTE YOUR TEXT HERE]

---

The prompt works on full articles, individual paragraphs, or single sentences. If you are running it through the ChatGPT interface, paste it as a new message after the draft appears. If you are using the API, you can include it as a user-turn instruction after the assistant's response.

The instruction about semicolons matters. A naive fix to "remove em dashes" sometimes swaps in semicolons at the same rate, which trades one tell for another. The prompt above forces the model to make a genuine structural choice rather than a character swap.

For a broader rewrite that addresses more AI patterns at once, see the full humanizer prompt and what it targets.

Before and After Example

Here is a short paragraph as a raw AI draft, followed by the same paragraph after running the prompt above.

Before:

Our platform helps teams move faster, smarter, and more confidently. The onboarding process is simple -- you sign up, connect your tools, and you are live in minutes. The dashboard gives you everything you need: real-time data, instant alerts, and full control.

(Note: the example above uses two hyphens in place of the em dash character to illustrate the pattern without reproducing the mark in this article's prose.)

After:

Our platform helps teams move faster without adding overhead. Onboarding takes about ten minutes: sign up, connect your tools, and you are live. The dashboard shows real-time data and surfaces alerts so you can act before a problem compounds.

The after version has no three-item lists. Each sentence carries a single clear idea. The rhythm is less even, which is actually the goal. Human writers do not naturally produce perfectly parallel structures at the same cadence throughout a document.

Adjusting the Prompt for Your Use Case

The base prompt above works for most general copy. A few adjustments are worth knowing about.

For academic writing: Add a line that says "Avoid hedging phrases like 'it is important to note' or 'this suggests that' used more than once per page." Academic AI output often pairs em dashes with these soft qualifiers, so targeting both at once saves a second pass.

For marketing copy: Add "Do not use the word 'transform' or 'revolutionize.' Prefer concrete verbs that describe what the product actually does." This pairs well with the rule-of-three instruction because marketing drafts tend to cluster these patterns together.

For social media: Shorten the prompt significantly. Most social copy is already short enough that you can handle the em dash issue with a single instruction: "Remove any em dash and rewrite the clause as two shorter sentences or a comma construction."

If you want a prompt tailored to sound like your own voice rather than generic edited prose, this approach to writing ChatGPT prompts that match your style covers that in more detail.

Where This Fits in a Full Editing Workflow

Removing em dashes and three-item lists is a surface-level fix. It matters, and it takes seconds, but it is one step in a larger process. The deeper issues in AI copy tend to be structural: paragraphs that explain rather than show, transitions that announce what the next sentence will say, and a tendency to conclude by restating the opening.

A good order of operations is to run structural edits first (cut filler, tighten the argument) and then run a targeted pass like this prompt to catch punctuation and pattern issues that survive the first round.

If you want a more systematic approach to building a system prompt that handles multiple tells at once, this guide to writing a system prompt that strips AI tells lays out the full process.

The prompt on this page also fits naturally into the workflow described at /humanizer-prompt, where the broader humanizer handles voice, filler phrases, and structure. Think of this em dash and rule-of-three prompt as a fine-tuning pass you run after the main humanizer has done its work.

Frequently Asked Questions

Does removing em dashes actually help with AI detection?

It can, depending on the detector. Some tools flag em dash density as a signal because it is statistically higher in AI output. Reducing that frequency lowers one signal. It does not guarantee a clean score because detectors look at many signals at once, but it is a real and measurable improvement.

Can I just do a find-and-replace on the em dash character instead?

You can, and it will remove the character. The problem is that the underlying sentence structure stays the same, so you get a period or comma where an em dash was without any rethinking of the clause. A rewrite instruction produces cleaner results because the model reconsiders the sentence rather than just swapping punctuation.

Does this prompt work in Claude, Gemini, and other models?

Yes. The instruction is plain English and model-agnostic. You may need to adjust the tone slightly for different interfaces, but the core logic applies to any instruction-following language model.

What if the model ignores the em dash rule and still uses them?

Add the phrase "This is a hard rule, not a preference" to the end of rule 1. Most models respond to that kind of explicit constraint. If the output still contains em dashes, run a second pass with just that instruction and the offending sentences.

How many em dashes is too many in a finished piece?

There is no universal rule, but one or two in a 600-word article is generally unnoticeable. Three starts to register. Five or more in that length is where readers sense something is off, even if they cannot name the specific mark. Targeting zero for AI-generated drafts is a reasonable default because it is easier to add one back deliberately than to remove several after the fact.