Custom instructions that stop em dashes and make AI writing sound human

They apply to every new chat in that account. They do not apply to a shared conversation someone else started, and they do not follow you into a different product, so set it wherever you actually write.

Short, explicit, and it names the substitutes. This is the block to use if you change nothing else.

Do not use em dashes (—), en dashes (–), or typed double hyphens (--). Use commas, colons, semicolons, parentheses or full stops instead. Keep the sentence structure otherwise unchanged when you make the substitution.

The last sentence matters more than it looks. Without it, models sometimes avoid the dash by splitting one sentence into three, which trades a punctuation tell for a rhythm tell.

Removing dashes without changing anything else still leaves the shape readers recognise: uniform sentences, no specifics, a summary at the end. This block addresses those directly.

Write in plain English. Never use these words: delve, tapestry, testament, realm, landscape, leverage, seamless, robust, meticulous, vibrant, crucial, pivotal, navigate, underscore, foster, holistic, myriad, comprehensive, streamline, elevate. Do not open with a restatement of my question. Vary sentence length deliberately and include at least one sentence under six words. Do not end with a summary paragraph. Use contractions. Prefer a specific number, date or name over an adjective wherever you can.

That is roughly 90 words, which is close to the practical ceiling for style rules that keep working. If you want more, put the overflow in a project-level instruction or a saved prompt instead of the global field.

• They do not survive a different account, a browser you are not logged into, or a shared workspace.
• They lose to an explicit in-chat request. Ask for "punchy marketing copy" and the banned register comes back.
• They cannot add facts. Style rules make generated text less obvious; they do not make it true.
• They do not apply to uploads or code blocks, so pasted-in content keeps whatever punctuation it arrived with.

The practical consequence: for anything you are going to publish or submit, finish with a mechanical pass. The AI dash remover takes a paste, removes the dash family, and shows a count so you can see what the instructions missed.

1. Open a new chat and ask for 300 words on any topic. Do not mention style.
2. Count the dashes with the em dash counter. Under one per 300 words means the block is working.
3. Now ask for "engaging, punchy copy" and count again. This is the stress test, and it is where most setups fail.
4. If it fails the stress test, add the character ban to the project instructions for that chat, or just clean the output.