How to Create a Style Guide Your AI Knowledge Base Can Use

Your brand voice shouldn’t feel like a costume change from channel to channel—friendly on the site, stiff in emails, and mysteriously emoji-happy on social. That inconsistency chips away at trust and slows teams down.

It also hits the bottom line: companies that present their brand consistently can see revenue lift by as much as 33%. Who wouldn’t want that?

As more of your content is drafted, summarized, or QA’d by AI assistants, the rules you give those tools matter just as much as the rules you give your team.

The goal is simple—create a style guide that eliminates debate, travels with your work, and plugs neatly into the knowledge features of tools like ChatGPT (via custom GPTs), Notion Q&A, Confluence AI, and your support platforms.

This article walks through the essentials you need to nail: the exact components of an AI-ready style guide, example rules you can copy, and a lightweight “LLM Appendix” that makes your guidance machine-friendly.

Then it shows how to upload and maintain that guide inside popular AI knowledge bases so your website, blog, social posts, emails, sales decks, and support replies all sound like the same smart brand—every time.

why a style guide matters (especially with AI in the mix)

consistency drives growth

Brands that show up the same way across channels can see revenue lift by up to 33%—proof that clear rules don’t just look tidy; they pay.

The point is, coherent style guide produces coherent customer experiences, which in turn drives trust and conversion. 

searching is expensive

Knowledge workers still burn roughly one day a week just hunting for information.

Centralizing plain-English rules in one place (and putting that document where people and your AI actually search) cuts that waste and speeds up shipping. 

AI amplifies your inputs

If your guide is vague, you’ll get vague answers—faster. Give the model unambiguous rules (e.g., “Use sentence case for H1–H3”) and short Do/Don’t examples, and you’ll see tighter, on-brand output.

This isn’t magic; retrieval systems perform best when they can pull small, semantically coherent chunks of guidance rather than wade through long, muddy paragraphs.

In other words, structure your rules so the machine can find and apply them quickly.

make your rules do the work

At the end of the day, a practical, example-rich style guide is a force multiplier—for humans who need quick decisions and for AI that needs unambiguous instructions.

Do that, and every page, email, and reply starts sounding like the same smart brand—consistently.

what makes a good AI-ready style guide (principles you can copy)

Here’s a concise set of rules to adopt. They’re built for both humans and AI:

• Be decisionable. Every rule resolves a choice (e.g., “Use the Oxford comma.”), not “it depends.”
• Show, don’t tell. Pair rules with short Do/Don’t examples—ideally lifted from your own content.
• Set precedence. Spell out what wins when rules conflict: Brand overrides → Channel overrides (web/blog/email) → Global rules → Fallback stylebook (e.g., AP).
• Scope per channel. Include tight appendices for web pages, blogs, newsletters—structure, length, formatting, and SEO norms.
• Make it machine-friendly. Add a one-pager LLM Appendix with binary choices (yes/no) and fixed options (enums), plus a glossary and exception list. Keep wording simple and specific.
• Include rationale & edge cases. A one-line “why” helps humans; explicit edge cases keep AI from drifting.
• Keep it maintainable. Put Version / Owner / Last updated at the top with a tiny changelog; keep the full guide under ~10 pages and link to deeper references.

Here’s a quick tip: before you hit publish, run drafts through a content quality checklist—H2/H3 structure, short paragraphs, links, and a final accessibility pass (contrast, alt text, descriptive buttons). It’s a small step that prevents big rewrites later.

The core sections to include (plain-text, copy-ready)

1. global voice & tone

• Voice: confident, practical, friendly; avoid hype.
• Readability target: Grade 8–10.
• Person & POV: second person (“you”).
• Inclusive language: avoid ableist/gendered idioms; use people-first phrasing.

2. mechanics (the “never-argue” rules)

• Headlines & headings: sentence case for H1–H3; proper nouns keep their case.
• Exceptions: product/UI strings keep their exact casing (e.g., iPhone, YouTube).
• Trademarks/brand marks: add ™/® on first mention per page/post/email; don’t verb or pluralize marks; keep official casing in a brand dictionary.
• Numbers: spell out one–nine; use numerals for 10+ and all units, dates, ages, dimensions, times, and percentages (use %).
• Dates & times: choose a standard (e.g., September 17, 2025; 12-hour with am/pm) and stick to it.
• Oxford comma: use it.
• Dashes: em dashes—no spaces—for emphasis; en dashes for ranges (10–15); hyphens for compound modifiers.
• Quotes/apostrophes: smart quotes and proper apostrophes (it’s).
• Capitalization of UI/feature names: mirror the product UI (Settings > Notifications).
• Links: descriptive anchor text (not “click here”); open in the same tab for editorial content unless policy dictates otherwise.
• Images: alt text required; describe the function or content.
• Emoji & exclamations: define where allowed (newsletters: sparingly; blogs/web: rarely).

3. lists (so your steps don’t step on each other)

numbered lists (steps/processes)

• Lead-in sentence ends with a colon.
• Each list item starts with a capital letter.
• If items are full sentences, end with periods. If fragments, no periods.
• Prefer 1. 2. 3. for steps; use (1), (2), (3) only inside paragraphs.

bulleted lists

• Keep parallel structure and consistent item length.
• Don’t mix sentences and fragments in the same list.

4. brand & terminology

• Brand dictionary: exact casing, spacing, symbols, and first-mention rules for company, product, feature, and partner names.
• Banned → preferred: utilize → use; AI tool → AI assistant; clients → customers.
• Acronyms: define at first use; maintain a list of “always define” vs. “common/okay without definition.”

5. Channel appendices 

Web pages

• One H1; H2/H3 for structure.
• Intro ≤ 50 words; paragraphs ≤ 90 words (scannable).
• Meta title 55–65 characters; meta description 150–160 characters; include primary keyword once.
• Minimum internal links; clear CTA placement; schema/image guidelines if applicable.

Blogs

• Working title vs. final H1; excerpt length; hero image ratios; byline/date format; categories/tags rules; pull-quote rules.
• SEO: primary and secondary keywords; header distribution; link policy to authoritative sources.

Newsletters

• Subject 45–60 characters; preview text 80–120 characters.
• One primary CTA; UTM required; footer/legal and unsubscribe compliance; image-to-text ratio; dark-mode checks.

6. accessibility & compliance

• Color contrast, link underlines, descriptive buttons, table captions, and no essential text baked into images.

7. workflow (so rules actually get used)

• Briefing checklist → draft → peer edit → final QA (links, accessibility, meta, spelling) → publish.
• Define approvers and SLAs. Include a short pre-publish QA list.
• Add an easy “propose a change” path to keep rules fresh.

8. fallback stylebook

• If not specified here, follow AP Stylebook (latest) and record explicit deviations (e.g., “Headlines use sentence case, not title case”).

Keep in mind: a simple, fill-in-the-blank worksheet helps teams adopt this faster than a dense PDF. Use your downloadable sheet to capture decisions, examples, and edge cases in one sitting.

Ready-to-use rule examples (drop-in)

headlines

• Do: “How to choose the right nozzle”
• Don’t: “How To Choose The Right Nozzle”

brand marks

• Do (first mention): “We recommend NimbusFlow® for workflow automation …” Then: “NimbusFlow features …”
• Don’t: “NimbusFlows” (don’t pluralize trademarks) or “to NimbusFlow your process” (don’t verb it)

numbered steps

Lead-in: “To update settings:”

1. Open the app
2. Tap Settings
3. Select Notifications.

em dashes

• Do: “This works—and scales—without extra tooling.”
• Don’t: “This works — and scales — …” (avoid spaces)

definitions

• Use a colon for definitions: “Three goals: speed, scale, quality.”

build your style guide step-by-step (plain text workflow)

1. Collect your source of truth. Gather voice/tone docs, mechanics, channel standards, accessibility rules, and any brand dictionaries.
2. Decide the non-negotiables. Lock the “never-argue” items: heading case, dashes, numbers, date format, Oxford comma.
3. Draft the LLM Appendix. Convert rules into yes/no decisions and fixed options (see template below). Keep sentences short and specific.
4. Add examples. For every rule, include a short Do/Don’t pair. Use your own content to make examples realistic.
5. Create the brand dictionary. Exact casing, marks, first-mention policy, banned → preferred terms, and acronyms.
6. Publish and version it. Put Version / Owner / Last updated at the top, plus a three-line changelog.
7. QA before publishing. Scan for structure (H2/H3), clear headings, short paragraphs, sufficient internal/external links, and a PEP-style intro and 4R outro.

the “LLM Appendix” (no code—just plain text)

Use this as a one-page appendix at the end of your guide. Keep it skim-friendly.

document metadata

• Version: [e.g., 1.3]
• Owner: [team/role]
• Last updated: [YYYY-MM-DD]
• Precedence: Brand overrides → Channel overrides → Global rules → Fallback stylebook (AP unless noted)

voice & tone

• Tone: confident, practical, friendly (avoid hype)
• Readability target: Grade 8–10
• Person/POV: second person (“you”)
• Inclusive language: avoid ableist/gendered idioms; people-first phrasing

mechanics (choose & lock)

• Headings: sentence case for H1–H3; proper nouns keep their case
• Oxford comma: use it
• Numbers: spell out one–nine; digits for 10+ and all units, dates, ages, dimensions, times, percentages (%)
• Dates: Month D, YYYY (example: September 17, 2025)
• Times: 12-hour with am/pm
• Dashes: em dash without spaces for emphasis; en dash for ranges; hyphen for compound modifiers
• Quotes/apostrophes: smart quotes; proper apostrophes (it’s)
• UI/feature names: mirror the product UI exactly (Settings > Notifications)
• Links: descriptive anchor text; same tab for editorial content
• Images: alt text required (describe function or content)
• Emoji: newsletters—sparingly; blogs/web—rarely

lists

• Numbered steps: lead-in ends with a colon; each item capitalized; sentences end with periods; fragments don’t
• Bulleted lists: keep parallel structure; don’t mix sentences and fragments

brand & terminology

• Brand dictionary: exact casing, spacing, symbols, first-mention rules (e.g., “NimbusFlow®” first mention; “NimbusFlow” thereafter)
• Banned → preferred: utilize → use; AI tool → AI assistant; clients → customers
• Acronyms: define at first use; maintain “always define” vs. “okay without definition”

channel add-ons (one line each)

Web pages: one H1; intros ≤ 50 words; paragraphs ≤ 90 words; target SEO meta conventions

Blogs: working title vs. final H1; hero image ratio; categories/tags rules; link to authoritative sources

Newsletters: subject 45–60 chars; preview text 80–120; one primary CTA; UTM required; dark-mode checks

accessibility & compliance

• Color contrast; link underlines; descriptive buttons; table captions; no essential text in images

workflow

• Brief → draft → peer edit → final QA (links, accessibility, meta, spelling) → publish
• Approvers and SLAs listed here
• Changelog: 1–2 bullets per update

fallback

• If not specified, follow AP Stylebook (latest). Deviations: [list here]

how to upload your style guide into popular AI knowledge bases

Think about it: the style guide only helps if it sits where your AI (and teammates) actually look. Here’s how to place it.

ChatGPT (custom GPTs with Knowledge)

• Create a custom GPT for your organization.

• In Knowledge, attach your “Style Guide” and “LLM Appendix” files (you can attach up to 20 files).

• In Instructions, add a plain-language rule: “Always apply our Style Guide and LLM Appendix before answering. When rewriting, cite the rule and provide a short Do/Don’t example.”

• Share the GPT across your workspace.

Quick formatting tip: Upload single-column, text-heavy versions (Doc/MD/PDF with selectable text) so the parser can ingest cleanly. OpenAI’s guidance confirms Knowledge is best for stable docs like style guides and explains how attached files are used during answers. 

Notion Q&A

• Place “Style Guide” and “LLM Appendix” as pages in a dedicated teamspace.
• Enable Q&A and include that team space as a source.
• Teach teammates to ask, “What’s our rule for em dashes?” or “Rewrite this paragraph in our voice.” Notion will answer using your pages.

Confluence + Atlassian Intelligence

• Publish your style guide as a Confluence page in a “Brand Ops” space.
• Ensure Atlassian Intelligence is enabled for Confluence Cloud.
• Create a short page titled “LLM Appendix” to keep atomic rules easy to retrieve.

Intercom Fin (Support AI)

• Put tone and phrasing rules in an internal Knowledge article.
• In Knowledge → Sources, make sure the style pages are enabled for Fin. Intercom notes that internal content is ingested almost instantly, while external URLs refresh weekly—so authoring natively keeps the AI current.

Jasper (Brand Voice + Memory)

• In Brand Voice, upload your style guide file or URL to teach Jasper your tone and rules; then refine the description.
• Use Memory to store product facts/terminology so Jasper references them consistently across outputs. Keep your brand dictionary here.

Perplexity (Files & Connectors)

• Upload your style guide and LLM appendix as Files (Perplexity supports text/PDF and more) so answers can reference them directly.
• For ongoing use, organize files and enable connectors if you want Perplexity to search a live repository in addition to uploads.

Bear in mind: Whatever the platform, the pattern is the same—name the files clearly, keep rules atomic, and store the appendix next to the main guide.

if you’re running a custom RAG stack (or planning to)

You don’t need to show code to get this right. But a few behind-the-scenes practices will make your KB sing:

• Chunk by structure, not by arbitrary length. Break documents into semantically coherent units (e.g., per heading/section) so retrieval pulls whole ideas, not sentence confetti.
• Enrich lightly. Clean headings, keep consistent terminology, and co-locate a rule with its example; these small steps improve semantic matching.
• Favor plain text and simple layouts. Avoid multi-column PDFs or heavy design that can confuse parsers.
• Evaluate retrieval, not just generation. If answers look off, inspect the chunks that were retrieved—odds are the problem is in the source, not the model.

governance: keep it fresh without making it a full-time job

• Assign an owner and set a quarterly review cadence.
• Keep a tiny changelog. One or two bullets per update at the top of the doc.
• Create a feedback loop. Add an internal “Propose a change” form; route requests to the owner.
• Watch AI for drift. If your agent answers off-brand, fix the source article; tools like Intercom update almost instantly when the content lives natively. 
• Onboard with a 30-minute workshop. “Using Our Style Guide with AI” — record it and store the link in the guide.

When all is said and done, the guide should feel short, sharp, and living—not a museum piece.

troubleshooting common issues (and quick fixes)

1. AI ignores tone rules

The Fix: Put the LLM Appendix in its own clearly named file/page and reference it explicitly in the tool’s instructions (e.g., “Apply ‘LLM Appendix’ before drafting.

Cite the rule used.”). OpenAI recommends giving explicit directions on how GPTs should use attached knowledge files. 

2. Old trademark guidance resurfaces

The Fix: Update the brand dictionary, then re-upload or republish inside your KB.

For Intercom Fin, content authored natively is ingested almost instantly; external URLs sync weekly—so keep brand rules as native articles where possible. 

3. Inconsistent date formats or list styles

The Fix: Convert narrative rules into yes/no decisions and fixed options, and keep a Do/Don’t example directly under each rule so retrieval returns them together.

This aligns with RAG guidance to provide semantically coherent chunks rather than long, mixed paragraphs. 

4. the AI “can’t find” your style guide in Notion

The Fix: It’s often a scope/permissions issue. Q&A only surfaces pages users can access and what’s in the selected sources.

Put the guide in a clearly scoped teamspace and ensure teammates have access. 

5. uploads parse poorly (weird spacing, missing sections)

The Fix: Avoid multi-column PDFs and heavy slide layouts. OpenAI notes the Knowledge parser works best with simple, single-column text; multi-column PDFs can cause extraction issues.

Re-export as single column or use DOC/MD.

6) Competing versions create drift.

The Fix: Establish a single Source of Truth with Version / Owner / Last updated at the top and a tiny changelog. Deprecate or archive older copies and link them to the current page.

Teams also see fewer retrieval mismatches when there’s exactly one canonical file in the Knowledge Base.

a short “using the guide” blurb you can paste into AI tools

Use this in the settings/instructions of your GPT, wiki AI, or support AI:

• “Always apply our Style Guide and LLM Appendix before answering or rewriting. Prioritize decisions in this order: Brand overrides → Channel overrides → Global rules → Fallback stylebook. When rewriting, name the rule you applied and include a short Do/Don’t example. If a rule is missing, ask for clarification.”

This little paragraph sets the paradigm and reduces ambiguity.

lock in your voice—everywhere

You’ve got what you need to move from “we should be consistent” to “we are consistent.” You learned why a style guide matters, what belongs in it, how to make it AI-ready, where to store it so tools actually use it, and how to keep it fresh.

In short, you’re now equipped to write a clear, compact guide, place it where work happens, and troubleshoot the usual drift.

When your rules are decision-ready and easy to retrieve, AI stops guessing and starts sounding like your brand—on the site, in emails, in support replies, everywhere. That means fewer rewrites, faster publishing, and a customer experience that actually feels cohesive. Who wouldn’t want that?

Next move: get hands-on. Join the AI Content Bootcamp to turn this into muscle memory—build your style guide, wire it into your tools, and pressure-test it with real prompts and use cases. 

If you remember one thing, let it be this: clear rules + real examples + smart placement = content your team (and your AI) can execute on today. See you in Bootcamp.