How to Build a Product Glossary AI Must Use in Every Draft

Write a glossary that gives each product, plan, feature and vendor name its exact form plus the wrong forms to avoid, load a short version where your team drafts (a ChatGPT or Claude project, a Gemini skill), then check every draft against the full list with a prompt or a find list. Update it the day anything is renamed.

The wrong forms matter as much as the right ones. Models learned product names from years of text, including every outdated and misspelt version, so telling one to use "Microsoft Entra ID" works better when you also tell it never to write "Azure AD". A glossary without its never-write column is only half a glossary.

Follow me on Instagram@sagnikteaches

Eight kinds of term that belong in a product glossary

A brand glossary for AI isn't a dictionary of your industry. It's the list of words your drafts keep getting wrong, plus the ones where being wrong would cost you. For an illustrative 12-person IT support firm, those fall into eight groups. Most businesses have all eight, even if the examples differ.

Connect on LinkedInSagnik Bhattacharya

Your own plan and product names

These are the names customers buy, so they must be exact. The firm sells two plans, Essentials Support and Complete Support (illustrative names), and AI drafts regularly produced "the Essentials plan", "Basic Support" and "Premium Support", a plan that doesn't exist. Each glossary row gives the exact name, the variants never to use, and one line on what the plan includes, because the AI will otherwise describe the plan from general knowledge of what IT plans usually include.

Subscribe on YouTube@codingliquids

Feature and service names inside the plans

Named features are where drafts blur most. The firm's "Priority Callout" is a specific paid option with an hourly rate; AI drafts turned it into "priority support", "emergency callouts" and "rapid response", each of which implies something slightly different. The glossary pins the name and adds a note: "Only available with Complete Support; never described as included."

Vendor product names

An IT firm writes about other companies' products constantly, and those names change. The glossary records the current name, the old names, and which one to use in customer-facing text. Microsoft renamed Azure Active Directory to Microsoft Entra ID in 2023; drafts still produce "Azure AD" because years of older material use it. In August 2026 Microsoft renamed the Microsoft 365 Copilot app to "Microsoft Copilot", while some product pages still say "Microsoft 365 Copilot Business" for the licence. Without a glossary rule, one proposal can use three names for the same thing.

Retired products and features

A separate list of things that no longer exist stops AI recommending them. For the IT firm: Microsoft Copilot Pro is retired (its consumer replacement is Microsoft 365 Premium); Microsoft Lens is retired, so scanning advice points to the scan feature in the OneDrive mobile app; ChatGPT agent was replaced by ChatGPT Work on 9 July 2026; OpenAI's custom GPTs stop running on 11 December 2026. Each row says what to write instead.

Terms with a house definition

Some ordinary words carry contractual meaning in your business. For the firm, "response time" means the time until a technician starts work, and "resolution time" is something it never promises. AI drafts treat the two as interchangeable, and "we resolve issues within an hour" in a proposal is a promise the contract doesn't make. The glossary defines the term in one sentence and bans the confusable one.

A translation agency has the same problem with its service names. "Certified translation", "notarised translation" and "sworn translation" are different services with different requirements, and a model will happily use them as synonyms. So are "revision" (comparing the translation with the source) and "proofreading" (checking the translated text alone). An agency glossary defines each in one line and states which the agency offers.

Words you never use

Some words are wrong for your brand even if they're correct English. The firm never says "unlimited support" (it isn't), "IT guys" (its team isn't all men, and the phrase undersells them) or "bulletproof security" (nothing is). These overlap with a general banned-phrase list, but the glossary holds the product-specific ones: the words that misdescribe what you sell.

Spelling and format conventions

Small inconsistencies make copy look careless: Wi-Fi or WiFi, email or e-mail, sign in or log in, MFA spelt out on first use or not. Pick one form for each and write it down with the variants. The firm's rule for multi-factor authentication is to spell it out once with "(MFA)" after it, then use MFA, and never to write "2FA" unless quoting a vendor.

Internal words that shouldn't reach customers

Every team has shorthand. The firm's technicians say "ticket", "P1" and "PSA" (its service-management software). Customers should read "support request" and "priority-one incident". AI picks up internal vocabulary from internal documents you give it, so the glossary maps each internal term to its customer-facing version.

Collecting candidates from documents you already have

You don't have to remember every term. The firm pulled candidates from six sources in about an hour: its price list, its standard contract and service levels, its website, three recent proposals, a month of support request summaries, and the notes in its AI error log. AI does the first pass well if you ask it to show variants rather than tidy them away:

From the documents below, list every product name, plan name, feature
name, vendor product name and technical term. For each one:
- show every different spelling or form that appears, with a count
- quote one sentence where it's used
Don't correct anything and don't merge variants. Group by type.
[price list] [contract] [three proposals] [website pages]

An illustrative extract of the output:

PLAN NAMES
  "Complete Support" (14)  "Complete plan" (5)  "Premium Support" (2)
  Example: "Premium Support includes unlimited on-site visits."
VENDOR PRODUCTS
  "Microsoft Entra ID" (3)  "Azure AD" (9)  "Azure Active Directory" (2)
  Example: "We'll configure conditional access in Azure AD."
TECHNICAL TERMS
  "response time" (11)  "resolution time" (4)
  Example: "Guaranteed 1-hour resolution time for P1 issues."

That extract found three problems before the glossary even existed. "Premium Support" appeared in two proposals, both with "unlimited on-site visits", a plan and a promise the firm doesn't offer. "Azure AD" outnumbered the current name three to one. And "1-hour resolution time" had crept into a proposal from somewhere. Each variant becomes a row in the never-write column.

The glossary table, filled in for an IT support firm

Keep the full glossary as a spreadsheet or a table in a document, with five columns. Twelve of the firm's 64 rows:

Three details make the table usable. The never-write column lists the exact wrong forms found in real drafts, not every conceivable variant. The note column carries the one fact that stops a wrong description, such as "No on-site visits included", because a correctly spelt plan name with a wrong description is still an error. And vendor rows say which name to use in customer-facing text, since customers often still search for the old one; the firm's help pages mention "(formerly Azure AD)" once, in brackets, for exactly that reason.

Plurals, possessives and names in the middle of sentences

Getting the spelling right is only half the job; product names also have to survive grammar. AI drafts bend names to fit sentences in ways a find list won't catch, and each one looks slightly amateur to a customer:

• Plurals of plan names. "Both Complete Supports include…" should be "Both Complete Support clients…" or "Complete Support includes…". A plan name is a name, not a countable noun.
• Possessives. "Essentials Support's response time" is clumsy; "the response time on Essentials Support" reads better.
• Lower-case brands at the start of a sentence. If a product is deliberately written in lower case, decide what happens at the start of a sentence, and give the rule an example. Most brands prefer rewording the sentence to capitalising the name.
• Shortened names in headlines. AI shortens long names when space is tight: "Entra" for Microsoft Entra ID, "Business Premium" for Microsoft 365 Business Premium. Decide which short forms are acceptable after first use, and list the rest as never-write.
• Names turned into verbs or adjectives. "We'll Complete-Support your office" is the extreme case; "Complete-Support-level service" is the common one.

The fix that works best is an example sentence for each tricky term, added as a fifth column. Models copy patterns more reliably than they follow abstract rules, so a row that says "Correct: Complete Support includes on-site visits. Wrong: Complete Supports include on-site visits." does more than a sentence explaining that plan names don't take plurals. The firm added example sentences to its 15 trickiest rows and put three of them into the short block as well.

A short version the AI reads every time

Sixty-four rows is the right size for a reference file and the wrong size for an instruction block. The longer the list of rules in front of a model, the more likely one gets missed, so the firm keeps a compact block of the 25 terms most often got wrong, with the most-violated at the top:

TERMINOLOGY RULES (always apply)
Plans: write "Essentials Support" and "Complete Support" exactly.
  Never: Premium Support, Basic Support, Complete plan, unlimited.
  Essentials Support has NO on-site visits.
Never promise resolution times. We promise response times only.
Microsoft names: Microsoft Entra ID (never Azure AD or AAD);
  Microsoft 365 Business Premium (never Office 365 Premium);
  Microsoft Copilot for the app, Microsoft 365 Copilot Business for
  the licence. Copilot Pro is retired: never recommend it.
Priority Callout is charged hourly and never "included".
Customer-facing text: "support request" not "ticket";
  "priority-one incident" not "P1".
Format: Wi-Fi; email; MFA spelt out once as multi-factor
  authentication (MFA).
If you're unsure of a product or plan name, write [TERM?] and
carry on. Don't guess.
The full glossary is in the file "Glossary v7".

The [TERM?] line is the most useful rule in the block. It gives the model a way out when it's unsure, and it turns a silent wrong guess into a visible flag the checker can resolve in seconds.

Where to load it so every draft sees it

A glossary only works if it's present when drafts are written, without anyone remembering to paste it. That means putting it in the tools' standing instructions, not in a document people are supposed to consult. As checked in September 2026:

• ChatGPT Projects. Available across ChatGPT plans, including Free. A project holds instructions and files, and chats inside it inherit both. Put the short block in the project instructions and upload the full glossary as a file. Projects can be shared on every plan, and ChatGPT Business adds group and workspace-link sharing, so the whole team drafts from the same rules; using ChatGPT Projects to keep client work separate shows the set-up, which is also how an agency keeps one glossary per client.
• Claude Projects. Available to all Claude users, with free accounts limited to five projects, but custom project instructions need a paid plan. On a paid plan, the project instructions take the short block and project knowledge takes the full glossary. On the free plan, add both as project files and open each chat by asking Claude to apply the terminology rules.
• Gemini. Google now offers skills, reusable instructions that Gemini can apply automatically when they're relevant, or that you call by name; up to 100 can be active at once. Gems are being moved over to skills, from November 2026 for personal accounts and March 2027 for Workspace business accounts, and Google says existing Gems will be transitioned automatically. A terminology skill that triggers on any drafting request is a good fit for a glossary.
• Not a new custom GPT. OpenAI's custom GPTs stop running on 11 December 2026, so don't build a new home for your glossary in one. If it currently lives in a GPT, move it to a project now; OpenAI's own migration turns a GPT's instructions into a skill and its knowledge files into reference files.
• Grammarly. For teams that write in many different apps, Grammarly Pro (list price $144 a year per member on annual billing, or $30 a month on monthly billing) includes one style guide, where you can add rules for how specific terms are spelt, capitalised and used. It then flags violations as people type, whether the text came from AI or not.
• DeepL, for translations. If drafts are translated, DeepL's glossary feature fixes how specific terms translate and adapts them grammatically in context; the number of glossaries and entries depends on the plan. Product names that must never be translated belong there as do-not-translate entries.

If a tool has nowhere to hold standing instructions, keep the short block at the top of a saved prompt in your team's shared prompt library, so it's pasted by default rather than from memory.

The same glossary for chatbots and voice agents

Drafts aren't the only place AI uses your terms. A website chatbot and an AI phone agent speak to customers directly, with nobody checking each answer, so the glossary matters even more there. Load the short block into the chatbot's instructions and the full glossary into its knowledge base, exactly as you would for a drafting project, and rerun your test questions whenever the glossary changes.

Voice adds one problem text doesn't have: pronunciation. An AI voice agent may read "MFA" as a word, stumble over an unusual product name, or say a plan name with the stress in the wrong place. The fix is a "say it as" column, filled in only for terms the agent gets wrong in testing: "MFA: say M, F, A", or a phonetic spelling for an awkward product name. At the firm, three test calls to the after-hours agent turned up two terms it mispronounced and one it replaced with an internal abbreviation it had picked up from the knowledge base. All three became glossary rows.

The note column earns its keep here too. A chatbot asked "does Essentials Support include site visits?" needs the note "No on-site visits included" to answer correctly; a correctly spelt plan name alone gives it nothing to stop a helpful yes.

Checking drafts against the glossary

Instructions reduce errors; they don't remove them. The firm checks every customer-facing draft in one of two ways, depending on length.

For longer pieces, such as proposals, help articles and newsletters, a checking prompt in a fresh chat with the full glossary attached:

Check the draft below against the attached glossary.
List every place where the draft:
- uses a form from the "Never write" column
- describes a plan or feature in a way that contradicts the note
- uses an internal term in customer-facing text
- contains [TERM?]
Quote the sentence, name the rule and give the corrected wording.
Don't comment on anything else.
[draft]

An illustrative result for a proposal draft:

1. "We'll migrate your users into Azure AD" - Never write: Azure AD.
   Fix: "into Microsoft Entra ID".
2. "Essentials Support includes quarterly on-site health checks"
   - contradicts note: Essentials Support has no on-site visits.
   Fix: remove, or recommend Complete Support.
3. "Tickets are answered within the hour" - internal term, and
   "answered" blurs response time. Fix: "We respond to support
   requests within four working hours" (Essentials Support).
4. "[TERM?] for device management" - glossary term: Microsoft Intune.

Item 2 is the kind of error that matters most and is hardest to spot by eye, because every word in the sentence is spelt correctly. A find list would never catch it. That's why the note column exists and why the checking prompt is told to compare descriptions with notes, not just spellings.

For short pieces, such as social posts and email replies, a find list is faster. Keep the never-write column as a plain list and search the draft for each item, or paste the draft into a spreadsheet next to the list and use a simple formula to flag matches. Twenty-five search terms take under a minute to check. Grammarly's style rules do the same job automatically for anyone who has it installed.

Testing whether the glossary works

Before rolling it out, test the glossary the same way you'd test any instruction. The firm took five typical drafting jobs (a proposal section, a help article, a client newsletter item, a support reply and a LinkedIn post) and ran each twice: once without the glossary, once in the project with the short block loaded. It counted glossary violations in each.

• Without the glossary: 19 violations across the five drafts, about 3.8 per draft. Seven were vendor names, five were plan names or descriptions, four were internal terms, three were format conventions.
• With the short block: 4 violations, about 0.8 per draft. Two were plan descriptions (the hardest category), one was "Azure AD" inside a quoted error message where it was arguably correct, and one was "WiFi".
• After the checking prompt: all four were flagged; three were fixed and one, the quoted error message, was left deliberately.

Two changes came out of the test. The plan-description rule moved to the top of the block, and a line was added to the glossary allowing old vendor names inside quoted error messages. Rerun the same five jobs after every significant change to the glossary or when you switch AI tools, and keep the counts, because a rising number is the earliest sign the rules have stopped working.

Renames, retirements and who updates the list

A glossary is out of date the moment a vendor renames something, and vendors rename things often. In the twelve months to September 2026 alone, the firm's glossary needed rows for the Microsoft Copilot app rename, the retirement of Copilot Pro and Microsoft Lens, and ChatGPT agent becoming ChatGPT Work. A translation agency's glossary would need Memsource recorded as Phrase TMS, a name that has been current since 2022 but still appears in older drafts and job descriptions.

Give the glossary an owner, and give the owner three triggers:

1. A change you make. A new plan, a price change or a renamed service updates the glossary on the same day, before the website.
2. A change a vendor makes. Whoever reads vendor announcements adds the row when a rename is announced, not when it takes effect.
3. An error that slipped through. Every wrong-term error found after publishing becomes a never-write entry, fed from your AI error log.

Version the file ("Glossary v7, 22 Sep 2026") and keep a one-line change log at the top, so anyone can see what changed and when. When the full glossary changes, check whether the short block needs the same change; the two drifting apart is the most common way a working glossary stops working.

What it takes to build and keep

The firm's glossary, from nothing to tested, took about five hours: an hour collecting candidates with the extraction prompt, two hours deciding correct forms and writing notes for 64 terms, an hour on the short block and loading it into the team project, and an hour on the five-job test. Upkeep runs at about 20 minutes a month, most of it adding vendor renames and reviewing never-write entries from the error log.

Against that, count what a wrong term costs. A proposal that sells "Premium Support with unlimited on-site visits" either loses the firm credibility when corrected or commits it to visits it doesn't price for. A help article that tells clients to open "Azure AD" sends them looking for a menu that no longer carries that name. Each is a small error with an outsized effect on how competent the firm looks, and each is exactly what a glossary with a never-write column prevents. For the wider rules on grammar, format and tone that sit alongside it, the house style guide your team and AI both follow is the companion piece, and how translators keep terminology consistent with AI covers the same discipline across languages.

Glossary questions teams ask

How long can the glossary get before the AI stops following it?

There's no fixed limit, but the always-on block works best short: the 20 to 40 terms most often got wrong, most-violated first. Keep the full table as a reference file and use it for checking. If a rule keeps being missed, move it to the top of the block rather than adding more words around it.

Should the glossary go in the instructions or in an uploaded file?

Both, for different jobs. Put the short block of rules in the instructions, where the model reads it with every request. Put the full table in a file, which the model can consult and your checking prompt can use. Test both with the same five prompts, because tools differ in how reliably they read uploaded files.

Can one glossary cover several clients or brands?

Keep one per brand. Shared terms, such as vendor product names, can live in a common file, but each client's plan names, banned words and spellings belong in that client's own project. Mixing them invites the worst kind of error: one client's product name appearing in another client's copy.

What about trademark symbols?

Decide once and write the decision into the glossary, because AI adds and drops symbols inconsistently. For vendor names, check the vendor's own trademark guidelines. For your own marks, ask whoever advises you on trademarks. Many businesses use no symbols in running text; whatever you choose, make it a rule with a never-write example.

How do we stop AI correcting a deliberately unusual spelling?

Add the corrected spelling to the never-write column, with a note saying the unusual form is intentional. Models tend to normalise odd capitalisation and joined-up names back to ordinary English, so a rule that names the exact wrong form they produce works far better than a general instruction to respect the brand.

Further reads

• How Agencies Keep Brand Voice Consistent in AI-Written Content — How agencies keep each client's voice consistent in AI content.
• How to Build a Brand Voice Guide That AI Can Follow — Cover tone and voice, which a glossary doesn't.
• How to Set Up Claude Projects as a Shared Team Assistant — Set up a Claude project the whole team drafts in.
• Brand Voice Tools Compared: Jasper, Writer and Custom Instructions — Compare paid tools that enforce terms and voice for you.
• AI Words to Avoid: A Banned-Phrase List for Your Brand — Build the banned-phrase list that sits beside the glossary.
• How to Check AI Translations Before Customers See Them — Check glossary terms survive translation into other languages.
• How to Check AI Is Doing Good Work, Not Just Fast Work — A rubric template, a 30-minute weekly sampling routine, a re-runnable test set and the warning signs, shown through a garden centre's plant-care replies.
• 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.
• 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.
• Can a Small Online Shop Use AI to Compete With Bigger Brands? — Compete through clearer product choices and dependable service, with a practical AI trial and an honest margin calculation.
• How to Use Gemini in Google Docs to Draft and Edit Faster — Drafting from @-mentioned files, Refine before and after, Help me create with a template, and the number mistakes to catch before you accept.
• How to Catch Outdated Information in AI Answers — How to catch stale facts in AI answers: the topics that date fastest, a prompt that exposes dates, a five-minute check and real 2026 examples.
• How to Audit Your Website for AI Content That Needs Fixing — A content audit for sites with AI-drafted pages: inventory, leftover searches, a five-check score, keep-fix-merge decisions and a translation agency's numbers.
• 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 (Projects in ChatGPT); Claude help pages (projects, free plan limits); Gemini Apps Help (skills, and the transition from Gems to skills); Grammarly Support (style guides and style rules, Grammarly Pro pricing); DeepL glossary feature page; Microsoft Learn (new name for Azure Active Directory); Phrase (Memsource renamed Phrase); our verified fact sheet (custom GPT retirement, Microsoft Copilot app rename, Copilot Pro retirement).