How to stop ChatGPT using em dashes

Two things conspire against a polite request. First, the instruction is a preference the model weighs against everything else it was asked for; when you also ask for "engaging" or "punchy" prose, that pressure pushes back toward the dash. Second, a model writing paragraph nine has no memory of the constraint as a running tally. It knows the rule, and it still chooses the dash locally, because that is the fluent token.

Paste this at the end of your request, after the task itself. It names the characters explicitly, gives a fallback, and closes the escape hatch of "heavier punctuation" rewrites.

Formatting rule for this reply: do not use em dashes (—), en dashes (–), or typed double hyphens (--). Where you would use one, use a comma, a colon, a semicolon, parentheses, or a full stop instead. Do not replace the dash with a longer sentence or a new sentence just to avoid it. Before you send the reply, scan it and remove any of these three characters.

Three details do the work: naming the characters rather than the concept, listing the permitted substitutes so the model has somewhere to go, and asking for a re-read. The last line is worth keeping even though it is only a partial safeguard.

No em dashes, no en dashes, no -- in the output. Use commas, colons or full stops instead.

• "Write naturally" — models think dashes are natural, because their training corpus is full of them.
• "Avoid fancy punctuation" — too vague. It will drop the dash and add an Oxford-comma pile instead.
• "Sound like a human" — pushes toward contractions and slang, and does nothing about the dash.
• "Don't be like AI" — mentions the word AI, which tends to produce a disclaimer about being an AI.
• "Use simple punctuation" — works sometimes, but leaves the en dash and -- untouched.

A per-chat instruction has to be repeated forever and is forgotten the moment you start a new thread. Custom instructions in ChatGPT, personal preferences in Claude, and custom instructions in Gemini all apply to every conversation and cost you one setup pass. The exact wording, and the companion instruction that makes AI writing sound less like AI, are in the custom instructions guide.

Even then, treat it as a strong default rather than a guarantee. Instructions change the base rate; they do not make the dash impossible. That is the honest reason the next section exists.

A prompt asks the model to behave differently on every token. A text transformation does not involve the model at all. That is the whole argument for cleaning the draft afterwards, and it is why the workflow below is what actually holds up across a long document, a long conversation, or a different model entirely.

1. Copy the finished reply.
2. Paste it into the AI dash remover, which is preconfigured for em dashes, en dashes and typed -- or ---.
3. Press remove. The counter tells you exactly how many it found, so you know what your prompt missed.
4. Copy the cleaned text back. Nothing else about the wording changes.

If the count is consistently low, your prompt is doing its job and the remover is catching the residue. If it is consistently high, skip the prompt and go straight to the remover — you will save tokens and get the same result.