Give your AI a house writing style instead of banning words one by one

Who this is for, and the job

You use AI to draft customer replies, posts or FAQ answers, and you keep fixing the same things. Too long. Too polite. "Rest assured." An English reply to a customer who wrote in Bahasa Malaysia. So you add another line to your prompt: don't say this, never say that. A week later the AI has a new favourite phrase, and the list keeps growing.

A list of banned words only tells the AI what not to do. It never says what to do instead.

The job here: write one short house style file, once, and put it wherever your AI keeps standing instructions. It says who reads your messages, how your business sounds, which words are yours, and what a good message looks like. Then every draft starts from your style instead of the AI's default voice.

What you need first

• 3 to 5 messages you or your staff sent that you were happy with. Remove names, phone numbers and addresses first (see the safe-to-paste check).
• 2 AI drafts you had to fix, and what you changed in them.
• The words you and your customers use for your services. Not the brochure words.
• A place for standing instructions: the instructions of a Claude Project, a ChatGPT Project or a Gemini Gem. For coding agents, use CLAUDE.md (Claude Code) or AGENTS.md (Codex and other agents). Menus move; check your app.
• About 30 minutes.

How it works

The file has four short parts:

1. Who reads and how we sound. Who your customers are, where they read, and a language rule: what happens when a customer writes in BM, English or a mix.
2. Plain-words rules. Six rules in your own words. Ours are built on the ideas George Orwell set out in his 1946 essay on plain English: prefer the short word, cut words that do nothing, say who does what, drop worn-out phrases, avoid jargon your reader would not use, and break any rule rather than write something awkward.
3. Your words. A small glossary of your terms, each with the word to avoid.
4. Good and bad examples. Two of each, from real (cleaned) messages. "Friendly" means different things to different businesses. An example shows the AI what it means for you.

Keep the whole file under one page. Anthropic's Claude Code docs say instructions files are read as context, not enforced settings, and that the more specific and concise they are, the more consistently Claude follows them. A short file you can read in one minute is also a file you will actually keep up to date.

What stays out: prices, opening hours and promises. Those are facts, and they belong in your price list or business files, not in a style file. Style decides how you say it; facts decide what you say.

The template (copy and fill in)

# House style: [business name]

## Who reads our messages
[Who your customers are.] Most read on [WhatsApp / email / Instagram], on [a phone / a laptop].

## How we sound
- Like [a friendly technician / a careful accountant / a helpful shop owner]: [3 plain words, e.g. polite, plain, direct].
- Language: reply in the customer's language. If they write in Bahasa Malaysia, reply fully in BM. If they write in English, reply in English. Do not switch languages inside one message. Our own terms below may stay as they are.
- "We" is the business. "You" is the customer.

## Plain-words rules
1. Pick the short, everyday word. "Start", not "commence". "Help", not "assist".
2. If a sentence still works without a word, take the word out.
3. Say who does what. "[Our technician checks the gas]", not "[the gas will be checked]".
4. Skip stock phrases (for example "rest assured", "do not hesitate"). Say the plain thing.
5. No trade words the customer would not use. If we need one, explain it once.
6. Break any rule above if following it makes the message unclear or rude.

## Format
- [WhatsApp: short lines, the answer first, one question at the end.]
- [Prices: "RM" + number, and say what is included.]
- [No exclamation marks or emoji unless the customer used them first.]
- [Never confirm a time, date or price that is not in our business files.]

## Our words
- [our term]: [what it means]. Not "[word to avoid]".
- [our term]: [what it means]. Not "[word to avoid]".
- [our term]: [what it means]. Ask for [detail] before quoting.

## Good examples
- "[a real message you liked, cleaned]"
- "[another one]"

## Bad examples (and why)
- "[a draft you had to fix]" ([why it is bad])
- "[another one]" ([why])

Set it up

1. Fill in the template. Use real messages for the examples, with names and numbers removed. If a rule does not fit your business, change it. The rules are a starting point, not a law.
2. Put it where the AI reads it every time. Paste it into your Project or Gem instructions, or save it as CLAUDE.md / AGENTS.md in a coding project. If you already have a business Project, add it as one section there (see Set up a Claude Project as your business brain).
3. Test it on three jobs you do every week, for example a WhatsApp reply, a Google Business post and an FAQ answer. Use a made-up customer. Run each once with the file and once in a plain chat without it, and read both.
4. Fix the file, not the draft. When a draft is wrong, ask why. Missing word? Add it to "Our words". Wrong tone? Swap in a better example. Resist adding another banned word unless it keeps coming back.

Worked example (test run on a made-up example)

We tested the file on a made-up aircond servicing business in Seremban, with made-up prices. One customer message, one deliberately generic "before" draft, one rewrite that follows the style file, then a short script to count patterns.

The made-up customer wrote: "Hi, how much for wash aircond? 2 units at home, Rasah Jaya. Can do this Saturday?"

Before (no style file, an excerpt):

Hi there! Thank you so much for reaching out to Kipas Biru Aircond! We truly appreciate your interest in our comprehensive aircond maintenance solutions. ... The pricing for this service will be determined based on various factors, including the size and condition of your unit. ...

After (following the style file and a made-up price note):

Hi, thanks for your message. We have two kinds of wash:
- General servis: RM60 per unit. We clean the filter and the indoor unit. About 45 minutes.
- Chemical wash: RM150 per unit for 1.0hp to 1.5hp, RM180 for 2.0hp. We take the indoor unit apart and wash it fully. About 1.5 hours.
For Saturday, I need to check with our technician before I can confirm a time.
Which one would you like, and what size are your 2 units? The size is on the sticker on the side of the indoor unit.

What the count showed on these two texts: 105 words before, 94 after, so only about 10% shorter. The real change was elsewhere: 0 prices before and 3 after, 4 exclamation marks and 2 emoji before and none after, no question before and one clear question after, and the Saturday slot no longer promised. The style file made the reply more useful, not just shorter.

We also checked the language rule with a made-up BM message ("Salam, berapa harga servis aircond 1.5hp?"). The reply stayed fully in BM, kept the two service names from the glossary, and ended with one question.

This is one run on one made-up example, written by us. It shows the file being followed. It does not prove how much better your drafts will get; your own test in step 3 does that.

Check before you use it

• Every rule is something you could point to in a draft and say "this breaks rule 3".
• The language rule says what happens when a customer writes in BM, English or a mix.
• Each glossary term has the word to avoid next to it.
• The examples are real messages, cleaned of names and numbers.
• No prices, dates or promises live in the style file.
• You ran the three-job test and the drafts with the file were better to you.
• A person still reads every customer message before it is sent (see Check before you send).

Watch-outs

• Style is not truth. The file shapes how a reply sounds. It cannot stop a wrong price or a promise you cannot keep. Keep facts in your price list (see Price list for AI).
• It will try to grow back into a ban list. If the file passes one page, cut. Replace three banned phrases with one better example.
• Instructions guide; they do not lock. The AI can still slip, especially in long chats. Start a new chat when replies drift.
• Mixed language is a choice. If your customers really write in Manglish and like replies the same way, write that into the rule on purpose. What hurts is the AI switching languages at random.
• No customer details in examples. The examples sit in your instructions for every chat. Clean them first.
• Use your own words for the rules. Orwell's essay is still under copyright in some countries, so we paraphrase his ideas rather than paste his list.

Related

• WhatsApp reply drafts: the style file makes those drafts sound like you.
• Set up a Claude Project as your business brain: where this file usually lives.
• Set up a project for Claude Code: for CLAUDE.md in coding work.

Credit and sources

• Inspired by Vox (@Voxyz_ai) and the earlier post that one quotes, on giving AI a writing system instead of banning words one at a time. Our template and rules are written from scratch.
• Matt Pocock (@mattpocockuk) on asking AI for simplified English and a shared vocabulary file. Our glossary section follows that idea.
• George Orwell, "Politics and the English Language" (1946), at the Orwell Foundation. Paraphrased above, not quoted.
• ASD-STE100 Simplified Technical English, a published writing standard for technical manuals, if you want a stricter rule set.
• Claude Code docs on instruction files: code.claude.com/docs/en/memory, checked 2 October 2026.