How to Write a House Style Guide Your Team and AI Both Follow

Write it as a short list of testable rules, each with a right and a wrong example, grouped by the decisions writers actually face: spelling, numbers, dates, names, formatting and required wording. Keep the AI version under about 5,000 characters, load it into a shared project your team drafts in, and test it on five real drafts.

Most style guides fail with AI for the same reason they fail with new staff: they describe preferences ("a clean, modern style") instead of decisions ("29 September 2026, never 29th September"). A person can ask what you meant. A model fills the gap with its defaults, which are the most common forms on the internet, not yours.

Follow me on Instagram@sagnikteaches

House style versus brand voice: which rules go where

A house style guide and a brand voice guide are different documents, and mixing them weakens both. Style rules have a right answer you can check with a find command; voice rules involve judgement and degrees. Keep them apart so the AI knows which instructions are absolute.

Connect on LinkedInSagnik Bhattacharya

The voice side, with how to describe tone so a model can imitate it, has its own tutorial on writing a voice guide for AI tools. Product and service names often grow into a list long enough to deserve their own file; the tutorial on building a product glossary AI must use covers that.

Subscribe on YouTube@codingliquids

The eight sections most small-business style guides need

You don't need a publisher's style manual. Two pages covering these eight areas settle almost every argument a small team has:

1. Spelling: your spelling variant, plus the handful of words your sector gets wrong.
2. Numbers and money: when to use words or figures, how to write amounts and percentages.
3. Dates and times: one format each, with the forms you don't use listed.
4. Names: your business name, service names, job titles, and what you call your customers.
5. Capitals and abbreviations: headings, subject lines, acronyms on first use.
6. Formatting: paragraph length, lists, bold, links, for each document type.
7. Required and banned wording: disclaimers, risk warnings, phrases compliance has ruled out.
8. Correspondence: greetings, sign-offs, subject lines.

Sections 7 and 8 matter most for AI, because they are where default AI output is most likely to be wrong for your business and where a mistake costs more than an awkward read.

Find your real rules in ten recent documents

Don't write the guide from memory. Your rules already exist, inconsistently, in what the team has sent. Gather ten recent documents (client emails, a report, a web page, a proposal) and list every place the same thing is written in different ways. A chat assistant speeds up the first pass:

Below are ten documents from our firm. List every item that
is written in more than one way across them: dates, times,
money, percentages, service names, job titles, how we refer
to clients, greetings and sign-offs. For each, show every
form you found and roughly how often. Do not recommend
anything.

[paste documents, with client names removed]

The planning firm's audit, after the owner checked the assistant's list with a find in each document (its counts were approximate, and it missed "yearly check-in" entirely), looked like this:

The audit took about 75 minutes, and each row became a rule. Two things make the decisions stick: pick the form already used most often unless there's a reason not to (less retraining), and decide quickly. The point is one answer, not the perfect answer.

Writing rules a model can apply

A rule a model applies reliably has three parts: the rule, an example of right and wrong, and any exception spelled out. Rewriting a typical guide's rules into that shape is most of the work:

Two further habits help. Put the rules your drafts break most often first, because instructions near the top of a long block tend to be applied more consistently. And phrase every rule positively with the wrong form as an example, not as a bare prohibition: "Adviser, never advisor" teaches both the target and the thing to avoid.

A filled-in guide for a four-adviser financial planning firm

Here is a complete AI rules block for an illustrative financial planning firm with four planners and two administrators. It is about 2,000 characters, comfortably inside the 5,000-character limit ChatGPT allows for custom instructions on paid plans, and short enough for any project's instructions field.

HOUSE STYLE v1.3 (updated 1 September 2026)
Apply to every draft. If a request conflicts with a rule,
follow the rule and say which one. Examples in attached
files may use old formats: these rules win.

1. SPELLING
- -ise, -our and -re spellings: organise, colour, centre,
  analyse. Programme for a series of events.
- Adviser, never advisor.

2. NUMBERS AND MONEY
- Words for one to nine, figures from 10.
  Exception: always figures for money, %, ages, tables.
- $1,250 (not $1250.00). $2.5 million in text.
- 4.5% (not 4.5 percent).

3. DATES AND TIMES
- 29 September 2026. Never 29th, never numeric dates.
- 9.30am, 2pm. Never 09:30 or 2:00 PM.

4. NAMES
- First mention: [Firm name]. After that: "we".
- Services, exactly: Financial Planning Review,
  Retirement Income Plan, Cashflow Forecast.
- Job titles lower case in sentences: "your financial
  planner, [first name]".
- They are "clients", never "customers".

5. CAPITALS AND ABBREVIATIONS
- Sentence case for headings and subject lines.
- Spell out on first use: exchange-traded fund (ETF).
- Write "for example", not "eg" or "e.g.".

6. FORMATTING
- Emails: no headings; paragraphs of three sentences max.
- Reports: section headings; lists of six items max;
  bold only for deadlines.
- Link text says where it goes. Never "click here".

7. REQUIRED AND BANNED WORDING
- After any mention of investment performance, add,
  word for word: "The value of investments can fall as
  well as rise, and you may get back less than you
  invested."
- Never: guarantee(d), risk-free, safe investment,
  beat the market, no-brainer.
- Never predict returns or state a future figure.

8. CORRESPONDENCE
- Existing clients: "Dear [first name],"
  First contact: "Dear [title] [surname],"
- Sign-off: "Kind regards," then the planner's full name
  and job title on separate lines.
- Subject: topic plus action and date, for example
  "Your Financial Planning Review: documents by 14 October"

Three details in that block are easy to miss and do a lot of work. The version number and date at the top let you tell which rules produced a draft. The line about attached files stops old documents overruling new rules (more on that below). And the required wording is in quotation marks with "word for word", which is the difference between the model reproducing a compliance-approved sentence and paraphrasing it into something nobody approved. If your firm is regulated, where AI fits at all is a question to settle first; what compliance allows for financial advisers using ChatGPT covers that ground.

Two versions from one source: the team guide and the AI block

Your team needs more than the AI does: the reason behind a rule, extra examples, which template to use for which letter. The AI needs less: rules and examples, nothing else, in a block short enough to be applied every time. Keep one master document for people and derive the AI block from it, never the other way round.

A chat assistant can do the first compression. A prompt that works:

Below is our 11-page house style guide for staff. Produce an
AI rules block under 4,500 characters:
- Numbered sections, each rule with one right and one wrong
  example.
- No explanations or reasons.
- Copy every piece of required wording exactly, in quotes.
- At the end, list every rule you left out.

[paste guide]

The planning firm's result, in part (illustrative):

7. REQUIRED WORDING
- After mentioning performance, add a risk warning:
  "Investments can go down as well as up."
...
RULES LEFT OUT: letterhead use; font sizes; table rules.

Two things needed correcting. Despite the instruction, the model shortened the approved risk warning into a common paraphrase, which is exactly the failure the block exists to prevent; restore the exact sentence. And "table rules" contained a real exception (figures for everything in tables) that belongs back in section 2. Letterhead and font sizes were rightly dropped: formatting the AI can't control belongs in your document templates, not in its instructions. Always read the "left out" list; it is where useful exceptions go missing.

Loading the guide where the team drafts

A rules block only works if it's attached to every draft automatically. Relying on people to paste it in fails within a week. Where to put it depends on your tools:

• Claude Team or Enterprise: put the block in a shared project's instructions, and consider an organisation skill, which owners can provision for everyone and which is switched on for all members by default.
• ChatGPT: a shared project with the block as its instructions. Projects can be shared on any ChatGPT plan; on Business you can also share with a group or with everyone in the workspace. Avoid building a custom GPT for it: OpenAI is retiring them, and they stop running on 11 December 2026.
• Google Workspace: a Gemini Gem with the block as its instructions, or a Gemini skill if you are starting now. Google is turning Gems into skills (from November 2026 for personal accounts, by March 2027 for Workspace business accounts), and existing Gems are recreated as skills automatically. People you share a Gem with can see its instructions and files, which is fine for a style guide.
• Staff writing without AI: Grammarly Pro ($30 a month or $144 a year) includes one style guide, so the same rules can be flagged as people type.
• A sole trader on an individual plan: ChatGPT's personal custom instructions hold up to 5,000 characters on paid plans, enough for the whole block.

Whichever you choose, keep the human version in the same shared drive as your letter templates, and put the formatting rules the AI can't apply (letterhead, fonts, margins) into those templates. If you are weighing a dedicated brand tool against these built-in options, the comparison of Jasper, Writer and custom instructions covers when a paid tool enforces style rules better.

Testing the guide on five real drafts

Don't assume the block works because it reads well. Run five briefs that force style decisions (dates, money, service names, a performance mention, a sign-off) and mark every breach. One illustrative test brief for the planning firm, run before version 1.3 existed:

Write a short email to [first name] confirming their annual
review on 14 October at 2pm and asking them to bring their
latest pension statement.

The illustrative output, with a proposed rule in place that said only "use our house style":

Dear [first name],

I'm writing to confirm your annual financial planning review
on Tuesday, October 14th at 2:00 PM. Please bring your latest
pension statement so your advisor can review your
portfolio's performance, which has been strong this year.

Kind regards

Six breaches in four lines: the service name isn't capitalised as "Financial Planning Review"; the date uses a day name and an ordinal; the time uses the wrong format; "advisor" is the wrong spelling; "which has been strong this year" mentions performance, makes a claim, and omits the risk warning; and the sign-off has no name or job title. With version 1.3 loaded, the same brief produced:

Dear [first name],

I'm writing to confirm your Financial Planning Review on
14 October at 2pm. Please bring your latest pension statement
so we can update your Cashflow Forecast before we meet.

Kind regards,
[Planner's full name]
Financial planner

Score each of the five test drafts by counting breaches, and keep the briefs: rerun them whenever the block changes or your AI tool updates its model. A block that produces zero or one breach per draft is ready; three or more means a rule is unclear or buried.

Where AI ignores a style rule, and what to do

When a rule keeps breaking, the cause is usually one of five things:

• Examples outvote rules. The planning firm's project also held 40 old client letters, all written with "October 14th". The model matched the letters, not the rule. Fix: replace examples with ones in the current style, or add the line telling it rules beat files, as version 1.3 does.
• The rule is buried. Rule 31 in a long block gets applied less reliably than rule 3. Move frequently broken rules to the top, and cut rules nobody breaks.
• Personal instructions compete. One planner's personal settings asked for an upbeat tone with exclamation marks. Ask staff to keep personal instructions free of style and tone rules for client work.
• The rule is about something the AI can't see. Fonts, logos and page layout belong in templates.
• Defaults creep back. Default spellings reappear in long drafts. Keep a short find list of your firm's usual slips, for example advisor, ize, October 1, PM, and run it on anything client-facing.

Style rules pair naturally with a list of phrases to avoid, which catches the generic wording no mechanical rule does; the tutorial on AI words to avoid gives a ready-made list.

A law firm's correspondence section, for comparison

The eight sections stay the same across businesses; the content changes. For an illustrative law firm, section 8 carries most of the weight, because client letters are the main output and small inconsistencies look careless in a legal context:

• Defined terms are capitalised consistently once defined: after "the lease dated 3 March 2024 (the Lease)", always "the Lease".
• Every deadline is stated with a date, never as "within 14 days" alone: "by 13 October 2026 (14 days from the date of this letter)".
• No Latin where plain words exist: "for example", not "e.g."; "in person", not "in propria persona".
• "We" means the firm; a named lawyer signs as themselves, with their role.

The deadline rule is the one worth copying into any business's guide. An AI draft that says "within 14 days" leaves the reader to work out which day counts, and that ambiguity is where complaints start.

Introducing the guide to the team

People follow a guide they have seen applied, not one they were emailed. Four steps take under an hour in total:

1. A one-page cheat sheet of the ten rules broken most often in the audit, for desks and the front of the shared drive.
2. A 15-minute walkthrough using the before-and-after email above. Six breaches in four lines makes the case better than any explanation.
3. The find list of your firm's usual slips (advisor, PM and the rest), so anyone can check their own client-facing draft in a minute before sending.
4. One route for disagreements. If someone thinks a rule is wrong, they propose a change for the changelog. They don't quietly override it in their personal AI settings, which is how five people end up with five house styles.

New starters get the guide and access to the shared project on their first day, before they draft anything a client will read.

Keeping the guide alive: owner, changelog and quarterly review

A style guide without an owner decays: rules get added by whoever was annoyed last, contradict each other, and the AI resolves contradictions by guessing. Give it one owner (for the planning firm, the operations manager), a changelog, and a quarterly review.

A filled-in changelog, which also becomes the record of why each rule exists:

At each quarterly review, rerun the five test briefs, read the last quarter's breaches, and apply two limits: every new rule needs a right-and-wrong example, and the AI block stays under 5,000 characters. When it nears the limit, something has to go, usually a rule nobody has broken in six months. The discipline of that limit is what keeps a guide short enough for both your team and your AI tools to actually follow.

House style guides and AI: other questions

Should we adopt a published style guide instead of writing our own?

You can start from one, but a small business rarely needs more than two pages of its own decisions. Published guides run to hundreds of rules, and an AI given all of them applies none reliably. Take the handful of decisions your writers actually argue about, add your names, required wording and correspondence habits, and stop there.

Can the AI check drafts against the style guide as well as follow it?

Yes, as a second pass: paste the draft and the rules block and ask for a list of rule breaks with the rule number, without rewriting. It catches most mechanical slips. It also misses some and invents a few, so treat its list as a prompt for a human check, and still run a find for the words your firm most often gets wrong.

How often should the style guide change?

Review it every quarter and change it whenever a service, price or required wording changes. Between reviews, collect proposed rules in the changelog rather than editing the live version. Frequent small edits by several people are how guides become contradictory, and a contradictory guide is one the AI resolves by guessing.

Further reads

• How to Write Reply Templates That Keep AI Replies On-Script — Reply templates that apply the style guide to customer messages.
• How to Build Gemini Gems for Repeat Business Tasks — The Google route for Workspace teams: Gems, now becoming skills.
• How to Set Up Claude Projects as a Shared Team Assistant — Setting up a shared Claude project to hold the rules.
• How to Check AI Content for Plain English and Readability — Readability checks that complement mechanical style rules.
• Grammarly vs ChatGPT for Checking Business Writing — Which checking tool enforces style rules better.
• How to Write an AI Content Policy for Your Marketing — The policy that sits above the style guide for marketing.
• How to Give AI Your Business Context So Answers Fit — Build a one-page business profile and task instructions, store them where each AI tool reads them, and test that answers fit your business.
• How Accountants Use AI to Explain Tax to Clients in Plain English — A facts-block method for turning tax computations into plain-English client explanations, with prompts, a readability check and a review routine.
• How Agencies Keep Brand Voice Consistent in AI-Written Content — A filled-in voice card, two client voices side by side, a sample-led prompt and a five-line scoring rubric for keeping AI copy on-brand.
• How Translators Use AI to Keep Terminology Consistent — A five-step terminology routine: AI term extraction, a client-approved termbase, glossaries in the engine, an LLM consistency audit and QA checks.
• Why Your AI Marketing Copy Sounds Generic (and How to Fix It) — Why AI defaults to the average of every ad it has read, the six fingerprints of generic copy, and the inputs and edits that make it sound like your business.
• A Five-Minute Fact-Check Routine for AI Output Before It Goes Out — A minute-by-minute fact-check routine for AI drafts, worked through on a price, a set of dates and a legal-sounding claim.
• How to Write AI Prompts That Sound Like Your Business — A copyable voice block for AI prompts, filled in for a barber shop, with before-and-after outputs and how to share one voice across the team.
• Should You Pay for an AI Humaniser Tool? What It Does and Risks — What AI humaniser tools really do to business copy, why beating detectors is the wrong goal, and the editing routine that fixes robotic drafts instead.
• Customer Service QA Software With AI: What to Compare — The criteria that matter when choosing AI quality assurance software for a support team, five products compared, and a worked choice for six agents.
• How Much Does Human Editing of AI Content Cost? — What editors charge to fix AI drafts, from a quick proofread to a full fact-check, with a costed monthly budget you can adapt.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI help pages and release notes (custom instructions character limits); Claude help pages (projects and organisation skills). Checked September 2026.