AI Writing Style Guide: Rules an AI Writer Actually Follows

An AI writing style guide is a short set of concrete, checkable rules, plus a few facts only your business knows, that you give an AI writer before it drafts anything. Keep it to a page or less. It works better than "friendly but professional" because an AI can follow "no em dashes" or "address the reader as you," but it has no idea what to do with "sound human."

This page covers the five kinds of rules worth writing, what to leave out, a copyable example for a fictional bakery, and a simple way to test your rules. One expectation to set first: good rules cut your editing time. They don't remove it.

Why adjectives don't change what an AI writes

Most brand voice documents start with tone words: warm, confident, clear. Hand those to an AI and you'll get roughly the same draft you got without them. "Confident" doesn't tell the model which sentence to cut or which word to avoid, so it falls back on its defaults.

Harvard's IT team puts it bluntly: generic prompts will produce generic results. The flip side shows up in a lot of practical advice too. One writer who drafts with AI sums it up as "More context = Better outputs".

A common tip is to list three to five adjectives that describe your ideal tone. That's fine as a summary for the humans on your team. For the AI, turn each adjective into something it can actually do:

The test for every line in your guide: could you look at a draft and say yes or no, it followed this? If not, rewrite the line until you can.

Step 1: Ban the words and phrases that keep showing up

Every AI model has favorite words. You've probably seen them: delve, game-changer, unlock, seamless, elevate, navigate, landscape. Rather than copying someone else's giant list, build yours from your own drafts.

1. Pull up your last several AI drafts.
2. Highlight the words and phrases that appear again and again, especially ones you'd never say out loud to a customer.
3. Keep only the ones that actually show up. Aim for 10 to 20.
4. Write them as a single rule: Never use: delve, game-changer, unlock, seamless, and so on.

Then add openings and closings, because those are where generic text clusters. Ban openers like "In today's fast-paced world" and "Let's dive in." Ban the closing paragraph that summarizes everything the reader just read.

Why not 50 words? A long banned list takes up space and attention that should go to your content rules, the ones about voice, examples and facts. It also tends to push the model toward awkward substitutes. A short list of words that really appear in your drafts does the job.

Step 2: Set rules for dashes, punctuation and formatting

Mechanics are the easiest rules to write and the easiest to check. You can scan a draft in ten seconds and see whether it broke one. Here's wording you can paste as is and adjust:

• Dashes: "No em dashes or en dashes. Use commas, periods, colons or parentheses instead."
• Headings: "Use headings only to start a new section. No headings inside short answers."
• Lists: "Use bullet lists only for steps or for items the reader will compare. Everything else is paragraphs."
• Length: "Sentences under 25 words where possible. Paragraphs of one to three sentences."
• Bold: "Bold at most one phrase per section, only for something a skimming reader must not miss."
• Emoji: "No emoji."
• Spelling: "Use US spelling and US date format (October 12, 2026)."
• Numbers: "Use numerals for prices, times, quantities and anything above nine."

Pick the ones that matter to you. If you don't care about emoji, leave that rule out.

Step 3: Turn your brand voice into rules

This is where brand voice guidelines for AI usually go wrong: they describe a personality instead of giving instructions. Swap the description for decisions.

• Who speaks: "Write as we, the team at [business name]."
• Who's addressed: "Address the reader as you."
• Reading level: "Write so a busy homeowner with no industry background understands every sentence on the first read."
• What you say plainly: "Give prices and timelines directly when we've listed them below."
• What you never say: "Never criticize other businesses by name. Never promise same-day service."

Then add the rule that does the most for specificity: "Include one example from our own work in each section." A plumber's post about water heaters reads very differently when every section mentions a real kind of job the business handles. Without that rule, the AI writes about water heaters in general.

One more trick: instead of explaining your voice, show it. Paste one short paragraph you wrote yourself and label it "Sample of our voice." A model matches an example far more reliably than it interprets a description.

Step 4: Add rules about claims and facts

This is the part people skip, and it's the one that can get you in trouble. The Authors Guild's guidance for writers states plainly that AI outputs cannot be trusted, and that AI tools often generate incorrect or fabricated information and present it as factual. Harvard IT adds that AI-generated content can be inaccurate, generic, biased, or contain copyrighted information.

You can't rule that risk away entirely, but you can make breaking a rule easy to spot. Phrase each rule so you can check it:

• "Never claim results, awards or rankings we can't back up."
• "No statistics unless I provide the number and its source in this prompt."
• "No customer stories, reviews or quotes unless I paste them in."
• "If you need a fact that isn't listed here, write [CHECK] instead of guessing."

That last rule is the useful one. It turns a silent invention into a visible gap you can fill.

Step 5: List the business facts only you know

The AI can't guess what makes your business yours. A short fact list is what turns a generic draft into one that could only be about you. Include:

• Location and service area
• Year you opened, or years in business
• What you sell, in your own words
• What you don't do (no delivery, no commercial jobs, no weekend appointments)
• Prices you're happy to state publicly
• Licenses, certifications or credentials you actually hold

The "what we don't do" line pulls more weight than it looks. It stops the AI from cheerfully promising services you don't offer.

A word of caution: by default, material entered into almost all public chatbots can be used to train those models. Keep anything confidential out of your rules, such as margins, supplier names or customer details. If you work inside a larger organization, check its AI policy and approved tools first, the way Harvard tells its staff to review the University's guidelines and use an approved tool.

What to leave out of your AI writing rules

Every extra line dilutes the lines that matter. Cut these:

• Vague wishes. "Sound human," "be engaging," "avoid AI tone." The model can't check itself against them, and neither can you.
• Contradictions. "Keep it short" next to "cover everything in depth." The AI will pick one at random.
• Huge banned lists. Fifty words crowd out your content rules.
• Brand history. Your founding story belongs on your About page, not in the instructions for every blog post.
• Things the AI already does well. "Use correct grammar" wastes a line.

A good limit: one page, or roughly 2,000 characters. If your guide runs longer, look for rules that repeat each other or can't be checked.

Example: a rules block for a local bakery

Example, fictional business. Fern & Flour is a made-up neighborhood bakery. Copy the block, then swap in your own details.

Voice: write as "we" (the Fern & Flour team) and address the reader as "you". Short sentences, plain words, no baking jargon without a one-line explanation.
Punctuation: no em dashes or en dashes. US spelling. Numerals for prices, times and quantities.
Never use: delve, game-changer, unlock, seamless, elevate, indulge, mouthwatering, nestled, "in today's fast-paced world". No closing summary paragraph.
Examples: each section includes one specific example from our bakery, such as a named bread or pastry, a step in how we make it, or a question customers ask at the counter.
Facts: Fern & Flour, Asheville, NC. Opened in 2015. Everything is baked fresh in our shop every morning. Pickup only, we don't deliver. We are not a certified gluten-free kitchen: never call any item gluten-free.
Claims: never claim awards, rankings or results we can't back up. No statistics unless I provide them with a source. No invented reviews, customer stories or quotes.
Format: paragraphs of 1 to 3 sentences. Headings only for new sections. Bullet lists only for ingredients or steps. No emoji. Bold at most one phrase per section.
Gaps: if you need a fact that isn't listed here, write [CHECK] instead of guessing.

Notice what's missing: no "warm and inviting tone," no history of the bakery, no list of 40 forbidden words. Every line can be checked against a draft.

How to test your rules

Don't trust a style guide until you've seen what it produces. Run this loop:

1. Generate a draft on a real topic, one you'd actually publish, not a throwaway test.
2. Read it aloud. Your ear catches stiff phrasing and repetition faster than your eyes do.
3. Run it through a patterns checker. Balzac's free AI writing patterns checker flags the recurring habits of machine-written text, which makes repeats easy to spot.
4. Write down what broke. Which rules did the draft ignore? Which patterns keep coming back?
5. Change one rule, then generate again on the same topic.

Change only one rule per round. If you rewrite three rules at once and the next draft improves, you won't know which change did it, and you might keep two rules that do nothing. A few rounds is usually enough to get a guide that holds up.

Where to put the rules

The best spot is wherever the AI will see them every single time, without you remembering to paste them.

• In a general chatbot: ChatGPT and Claude both let you save standing instructions, either as custom instructions or in a project with its own instructions or files. Put your rules there.
• At the top of every prompt: the low-tech option. It works anywhere, but it's easy to forget or to paste an old version.
• In your writing tool's settings: most AI writing tools have a field for brand voice or custom rules.

In Balzac, the rules go in Settings, Writing style, Writing rules (up to 2,000 characters), and every article Balzac plans, writes or rewrites for that site follows them.

What rules can't do

Even a great style guide won't produce text you can publish untouched. As one writer who drafts with AI puts it, "The AI-assisted content is not ready to copy, paste, and publish." Harvard's guidelines say all AI-generated content should be reviewed before publishing or sharing. Rules shrink that review from a rewrite to an edit, which is the real win.

A few other limits to know:

• No rule set makes AI text undetectable, and AI detectors are unreliable anyway. Don't aim to "pass" one. Aim for a draft that's specific, accurate and useful to your customer.
• Copyright is limited. According to the Authors Guild, AI-generated text is not copyrightable because it isn't original human authorship, and AI-generated text included in a work submitted for registration must be disclosed and disclaimed in the application (in the US, that's filed with the U.S. Copyright Office). Your edits and your own writing are what you can protect.
• Rules don't replace judgment. Even the Authors Guild, which first published its best practices for writers using AI in February 2024, frames that advice as guidelines, not rules. Treat your style guide the same way: a strong default you revise as you learn.

Your next step: open your last three AI drafts, build your banned list from them, and write the first version of your rules today. Then run one test round and change one thing.

Sources

• AI Basics | Harvard University Information Technology
• Create An Audience Style Guide For 10x Better AI Content
• Creating a Style Guide for On-Brand AI Content - Enrollify
• AI Best Practices for Authors