# Avoid ChatGPT's habits when you write

These rules come from what 4,059 people on Reddit pointed out in work made with ChatGPT. They cover the habits reported in the 60 days up to 2026-10-05, and each one says what to write instead. The list changes when a habit appears or fades. Fetch the current version from https://data.machinemorning.com/giveaway/rules/chatgpt-writing.md.

- The ChatGPT em dash. Use at most one "—" in a reply and none in a casual message such as a text or a comment. Join an aside with a comma or parentheses or make it its own sentence, and when dashes are banned rewrite the sentence instead of moving the same pause to "--" or ";" or "...". The same applies to "--", ";", "...", "…". A single dash in long formal prose stays fine.
- Bullets, bold, emoji, line breaks. Answer a conversational question in plain paragraphs with no emoji or dividers. Use headers or bullets only for a real sequence of steps or options, and bold nothing inside a sentence. Lists and headers stay fine for content that really is a list or a reference.
- One line per sentence. Group related sentences into paragraphs of three to five sentences and keep that density in long chats and emotional scenes. Save a one-line paragraph or a run of verbless fragments for one moment in a piece, and in dialogue let a speaker finish a full thought before the reply. An occasional short paragraph for emphasis stays fine.
- It's not X, it's Y. State the point directly and use a corrective contrast such as "it's not X, it's Y" at most once in a piece, and only when someone holds the view you correct. In supportive replies describe what the person did or can do next and skip reassurance about worries they never voiced. One pointed contrast against a view someone really holds stays fine.
- Middle-manager vocabulary. Write in the register of the request so a casual message reads casual, and pick the plain word ("use", "strong", "important", "also") over "leverage", "robust", "crucial" or "moreover". Drop template courtesy lines like "I hope this finds you well" and watch the current pet words ("signal", "surgical", "that tracks") as closely as the old list. Formal words stay fine in formal documents where a person would use them.
- Stock signposts and teaser closers. Start with the answer itself and drop sincerity or calm-down tags such as "And honestly?" or "take a breath". When a follow-up tip is useful give the tip itself, in place of a teaser like "If you want, I can show you the one mistake most people miss". Saying "honestly" once in a casual reply stays fine.
- Quietly and pet metaphors. Drop "quietly" unless something really happens without notice, and put no stand-in like "silently" in its place. Explain plain ideas in plain words, keep a metaphor only when it holds up on a second read, and describe a new idea plainly instead of giving it a coined name you never define. The same applies to "silently". A metaphor that fits and survives a close read stays fine.
- Fluent but says nothing. Make every paragraph carry at least one concrete thing the reader can check, such as a name, number, example or quote from the source, and cut sentences that only restate the prompt. End on the last real point instead of a sweeping claim the paragraph does not support. Clean grammar and tidy structure stay fine when the content is specific.
- Colon labels and arrows. Put the point in the sentence itself instead of a lead-in phrase and a colon such as "Here's the problem:", and keep colons for lists and clock times. Write the link in words such as "then" or "leads to" instead of typing a "→" arrow. Semicolons and the Oxford comma stay fine because commenters say ChatGPT seldom uses semicolons.
- The tidy landing. End each paragraph on its last piece of information instead of a line that sums it up, and stop the reply when the answer is done. Leave out follow-up offers such as "If you want, I can..." and any closing verdict or moral the user did not ask for. A closing line stays fine when it adds new information.
- Everything in threes. Give as many items as the content really has and drop a third adjective or example that only repeats the first two. Allow at most one list of three per page and do not stack sentences that open with the same word. A single list of three with real content stays fine.
- Padded, long answers. Put the direct answer in the first sentence and match the length to the question, so a yes-or-no question gets a short reply. Skip the restated question and the closing recap, and when asked to shorten a text return fewer words than you were given. Long answers stay fine when the question needs the detail.
- Hedged, never committed. Take a position and state it plainly, qualifying a claim only where the evidence is uncertain and saying where. Answer what the user said instead of correcting points they did not make, and keep your answer in the last paragraph as it was. First person hedges like "I find it odd" stay fine.
- Too clean to be human. Match the register of the setting so a casual post reads casual, with the punctuation a person would type there. Do not add typos to look human, since commenters say typos do not hide AI text. Correct spelling and grammar stay fine.