You asked for a LinkedIn post. You got “In today’s fast-paced world” and 3 em dashes before the first full stop.
Asking it to “sound more human” doesn’t fix that. A written list of rules does, saved once in the place your AI reads before every chat.
I’m not guessing here. The writing system behind this guide checks every draft against a list like this one. A draft with an em dash or a word from my banned list fails the check and gets rewritten before I read it. My personal rules sit on top: British spelling, short sentences, no role-play prompts.
Below you get the rules, the place to save them in ChatGPT and in Claude, and a check prompt for the drafts that still slip. If you want to go further with how you use AI day to day, that’s what my AI Power User workshop covers.
Which words and phrases make AI writing sound like AI?
AI writing sounds like AI because it reaches for the safe, important-sounding word instead of the specific one. Wikipedia’s editors keep a long list of these signs (Signs of AI writing), and my own list started there.
The giveaways fall into five groups.
1. Words nobody uses out loud
These are the words I ban first:
delve, crucial, robust, seamless, pivotal, tapestry, testament,
showcase, foster, elevate, unlock, game-changer, "the AI landscape"The important part: tell the AI to rewrite the sentence, not swap the word. A banned word swapped for a fancy synonym still sounds like AI.
2. Em dashes
The long dash is the most famous tell. Ban it, and say what to use instead: full stops, commas or brackets.
3. Sentence habits
Four patterns give it away:
It's not just X, it's Y.
Fast, simple and effective. (groups of three by reflex)
..., highlighting the importance of ... (a fake conclusion tacked on)
It's worth noting that ... (filler)4. Structure habits
Headers on a 200-word email. Bullet lists for everything. Bold on every second phrase. A summary paragraph repeating what you read 10 seconds ago. Then the sign-off:
In conclusion, ...
I hope this helps!5. Vague claims
“This can save businesses time” says nothing. The rule: use a number, a name or an example, and if the AI doesn’t have one, it asks you. That last part stops it inventing facts to sound specific.
Add your own layer on top: who you write for, your spelling, and 2 or 3 phrases you say. Leave out “You are a professional copywriter”. I explain why in stop adding roles to your AI prompts: audience, goal, format and constraints do the work a role pretends to do.
The full list is in the “My writing rules” prompt below, ready to paste. It’s a starting list, not a tested recipe, so trim it to what bothers you.
Where do you save writing rules in ChatGPT?
Save them in custom instructions, under Settings > Personalization on web and desktop (OpenAI help). They work on every plan, on web, desktop, iOS and Android.
Since November 2025, edits to custom instructions also apply to chats you already have open, not only new ones.
There’s a character limit, so keep the list tight. The “My writing rules” prompt is under 1,000 characters.
Two more settings on the same page help:
- Characteristics. Turn headers and lists down, and emoji down. These carry into every new chat (OpenAI help).
- Base style and tone. Efficient or Candid cut some of the cheerful padding. Pick by feel.
For one kind of writing, such as your newsletter, make a project and put the rules in its instructions. Project instructions apply only inside that project and win over your global ones (OpenAI help). Projects reached the Free plan in September 2025.
ChatGPT also has skills. OpenAI’s help page lists them for Business, Enterprise, Healthcare and Edu users. On a personal plan, stick with custom instructions and projects.
Where do you save writing rules in Claude?
Save rules you want in every chat under Settings > Your Claude > Instructions. Click your initials in the lower left to find Settings (Claude Help Center).
For one kind of writing, use a project: Set project instructions, paste your rules, save. They apply to every chat in that project. Projects work on every plan, and a free account gets up to 5 (Claude Help Center).
For a long rules list, make a skill. A skill is a folder with a SKILL.md file: a name, a description of when to use it, then your rules. You zip the folder, upload it, and switch it on in Customize > Skills (Claude Help Center). Claude loads it on its own when you write. Skills are on every plan, including Free, and need code execution turned on (Claude Help Center).
What about styles? Older guides tell you to save a custom style. Anthropic said styles were moving to skills, and that help page now returns an error. My own Claude account no longer shows a Styles option. Use Instructions, a project or a skill.
The “Make my writing rules a Claude skill” prompt turns your list into a SKILL.md for you.
Custom instructions, project or skill: which should you use?
Pick by how often the rules apply.
| Where | Best for | Applies to |
|---|---|---|
| ChatGPT custom instructions | Rules you want in every ChatGPT chat | All your chats, on every plan |
| Claude Instructions | Rules you want in every Claude chat | All your chats |
| A project (either tool) | One kind of writing, such as a weekly newsletter | Only chats inside that project |
| A Claude skill | A long rules list you don’t want loaded every time | Chats where Claude decides the task needs it |
A simple split: short personal rules in global instructions, longer rules per job. Once a job repeats every week, I stop prompting it and build a little app for it instead.
Why does the AI still break your rules?
Rules are instructions, not filters. The AI follows them most of the time. Long chats, long lists and requests that pull the other way (“make it exciting!”) all weaken them.
Three habits keep them working:
- Put the worst offenders first. The rules at the top get followed best.
- Keep the list short. Ten clear rules beat 40 vague ones.
- Check before you publish. Paste the draft and your rules into the “Check this draft” prompt. It lists each slip with the sentence and a fix, without rewriting your whole piece.
That last one is the habit I’d keep even if the first two worked perfectly. I don’t trust a draft until something has checked it.
Plans and settings checked on OpenAI’s and Anthropic’s help centres on 9 November 2026. For shortcuts people paste into ChatGPT, see ChatGPT secret codes and prompt shortcuts. Once your voice is clean, put it to work on same-day posts with newsjacking prompts for ChatGPT or Claude. For the difference between asking AI and building with it, read building with AI.
My writing rules (paste into custom instructions)
When you write for me:
- I write for [who reads my writing]. Use [British/US] spelling.
- Short sentences. One idea per sentence. Contractions are fine.
- No em dashes. Use full stops, commas or brackets.
- Never use: delve, crucial, robust, seamless, pivotal, tapestry, testament, showcase, foster, elevate, unlock, game-changer, landscape. Rewrite the sentence, don't swap in a synonym.
- No "not just X, but Y". No "it's worth noting". No "in conclusion".
- Don't group things in threes by habit. Use the real number.
- No headers, bullet lists or bold unless I ask.
- No summary paragraph at the end. Stop when the point is made.
- Be specific: a number, a name or an example beats a general claim. If you don't have the specific fact, ask me. Never invent it.
- Phrases I use: [2-3 phrases you say often].
Turn my writing into rules
Below are [2-3] pieces of my own writing. Read them and write a list of rules an AI should follow to write like me.
Cover: sentence length, words and phrases I use often, words I never use, how I open and end, punctuation habits, and formatting (headers, lists, bold).
Format: a plain bullet list under 1,400 characters, so it fits in ChatGPT custom instructions. Each rule is one line and specific enough to check. Only describe what's in my samples. Don't add rules from general writing advice.
My writing:
[paste samples]
Make my writing rules a Claude skill
Turn my writing rules below into a Claude skill.
Give me the full text of a SKILL.md file:
- It starts with a name (lowercase, hyphens, under 64 characters) and a description under 200 characters that says when to use the skill, for example "Use whenever writing posts, emails or articles for me."
- Then the rules, grouped under short headings: words, punctuation, sentences, structure, specifics.
- Then a short checklist to run before showing me a draft.
Keep my wording. Don't add new rules.
My rules:
[paste your rules]
Check this draft against my rules
Check the draft below against my writing rules. Don't rewrite the whole draft.
List every problem you find, one per line, with the sentence it's in and a suggested fix:
- banned words and phrases
- em dashes
- lists of three that don't need to be three
- filler like "it's worth noting" or "in conclusion"
- general claims with no number, name or example
- headers, bullets or bold I didn't ask for
If you find nothing in a category, skip it. End with the 3 changes that would help most.
My rules:
[paste rules]
Draft:
[paste draft]
Frequently asked questions
Which words make AI writing sound like AI?
The giveaways are grand, important-sounding words people rarely use out loud, plus filler openers and summary sign-offs. The guide lists the ones to ban first, and the My writing rules prompt holds the full list ready to paste. Tell the AI to rewrite the sentence instead of swapping in a synonym.
How do I stop ChatGPT using em dashes?
Put the rule in your custom instructions, not in a single chat. After GPT-5.1 launched in November 2025, Sam Altman said ChatGPT now follows a custom instruction to avoid em dashes, and press tests found it held up in custom instructions better than in one-off requests. Write the replacement too: use full stops, commas or brackets.
Should my writing rules go in custom instructions, a project or a skill?
Put rules you want everywhere in custom instructions or Claude's Instructions. Use a project when the rules belong to one kind of writing, such as newsletters, because project instructions only apply inside that project. Use a Claude skill when the rules are long, because Claude loads a skill only when the task needs it.
Can I save writing rules on the free plans?
Yes. ChatGPT custom instructions work on every plan, with a character limit, so keep your rules short. ChatGPT Projects reached the Free plan in September 2025. Claude Projects work on every plan, with up to 5 projects on a free account, and Claude skills are on the Free plan too.
Can I still use Claude styles for my writing rules?
Don't count on it. Anthropic said custom styles were moving to skills, and that help article now returns an error. I don't see a Styles option in my own Claude account any more, so save your rules in Instructions, a project or a skill instead.
Why does the AI still break my writing rules sometimes?
Because rules are instructions, not filters. Long chats, long rule lists and requests that pull the other way all weaken them. Keep the list short, put the most important rules first, and run a check prompt on any draft you're about to publish.