Keep a Glossary File Your AI Can Read

When an assistant keeps calling your thing by the wrong name, you do not need a fine-tune or a vector database. You need a glossary — a one-page plain-text file with each term, your definition, and what not to confuse it with. You paste it in as context, and it teaches the model your vocabulary for that conversation.

The fix most people reach for is bigger than the problem. They assume teaching a model their dialect means training data, embeddings, a pipeline. It rarely does. A glossary steers wording, not facts, and wording is most of what drifts. The grounding research is blunt about the mechanism: retrieving external knowledge, a survey of LLM hallucination notes, "can enhance the factual grounding, contextual relevance, and overall accuracy of LLMs."¹ A glossary is the smallest possible piece of that external knowledge.

What you keep re-typing every session

Terminology drift is the friction of correcting a model on your own words. You write "no, we call it a workspace, not a board," and next session you write it again. The model is guessing from a generic prior because nothing told it your definition. You are the missing context, supplied by hand, every time.

This is the shape worth naming. Not a bad app — a bad arrangement, where the meaning of your terms lives only in your head and gets re-explained on demand. Every guess the model makes about your domain is a guess it should not have to make. The cost is small per session and enormous across a year of them.

Why "just prompt better" stops working

Prompting better helps, until the vocabulary load grows. A sentence of correction works for one term. Ten terms, each with a near-synonym you reject, is no longer a sentence — it is a reference, and references belong in files, not in your memory of last Tuesday's chat.

The same survey frames the broader fix plainly: with retrieved context, "models can consult up-to-date and domain-specific data during generation, which in turn improves the factual accuracy of their outputs."² Domain-specific data during generation is the academic name for what a glossary does — the document you supply so the model is not improvising your terms.

Retrieval-augmented generation, the survey continues, "enhances LLMs by retrieving relevant external documents at inference time, which enables the LLM to ground its responses in factual evidence beyond its parametric training data."³ A glossary is the smallest such external document — no store, no retriever, no pipeline. You are the retriever. You paste the file.

The glossary, written once

Write the terms that actually drift, each in three parts: the term, your one-line definition, and the confusions to rule out. Keep it to a page. The format is boring on purpose. Boring is what a model parses cleanly, and what you can still read at a glance a year from now.

A worked example, for a team that keeps getting two words crossed:

# Project glossary — feed this in as context

Workspace
  Our definition: a single user's private container of notes and boards.
  Not to be confused with: "board" (a view inside a workspace) or
  "team space" (we do not have one — never invent it).

Board
  Our definition: one kanban-style view inside a workspace.
  Not to be confused with: "workspace" (the container) or "dashboard"
  (we have no dashboard; do not use that word).

Draft
  Our definition: a note not yet shared with anyone.
  Not to be confused with: "private note" (all notes are private by
  default; "draft" specifically means unshared-and-unfinished).

The third line of each entry is the part most glossaries omit, and the part that earns its keep. Not to be confused with is where you catch the model's most likely wrong guess before it makes it. Naming the near-miss is more useful than naming the term — because the near-miss is what actually drifts.

Which terms belong in the file is the only real judgment call. Include the words you have already corrected the model on, the in-house names that collide with a common meaning, and the synonyms you specifically reject. Leave out anything the model already gets right — every line you add competes with the rest of your prompt for the model's attention. The glossary is a list of fights you keep having, not a dictionary.

What a glossary fixes — and what it does not

A glossary fixes wording. It does not fix facts. This is the honest boundary — and skipping it is how a useful technique becomes a dishonest one. A model can produce a fluent, confident, wrong answer using every one of your correct terms.

The survey defines the thing a glossary cannot touch: a hallucination is "the generation of content by an LLM that is fluent and syntactically correct but factually inaccurate or unsupported by external evidence."⁴ The vocabulary was right; the claim was invented.

If you need the model to answer from your documents and cite them, that is retrieval — a personal RAG you can actually audit, not a term sheet. A glossary steers how the model says things. Retrieval governs whether what it says is true. They are different jobs, and the glossary only claims the first one.

Three more limits, stated before you hit them:

• It must stay short. A glossary competes for the context window, which IBM defines as "the amount of text, in tokens, that the model can consider or 'remember' at once."⁵ A bloated term sheet gets crowded out by the rest of your prompt. Keep it to the terms that drift; cut the ones the model already gets right.
• The model usually will not remember it. Paste once is not taught forever. Across a fresh session the definitions are gone unless you re-supply them, or keep them where they persist: inside the custom-instructions file your assistant reads. Your AI's instructions belong in a file you own, and a glossary is the vocabulary slice of that file.
• Even formal grounding is not a guarantee. A study on domain-grounded retrieval reports that its pipeline "consistently outperformed the zero-shot baseline across all tested environments, with win rates ranging from 50.0% to 83.7%."⁶ When a full retrieval system tops out near 84%, a one-page glossary is a floor, not a ceiling. It reduces drift. It does not abolish error.

A file, not a feature

Write your glossary as plain text for the reason you write anything worth keeping as plain text: the file outlasts the tool. A glossary locked inside one assistant's settings panel dies with that assistant. A glossary in a .md file moves to the next one untouched, which is the whole point of owning the words rather than renting them.

Steph Ango, who put the idea best, writes that "File over app is a philosophy: if you want to create digital artifacts that last, they must be files you can control, in formats that are easy to retrieve and read."⁷ A glossary in a settings panel breaks that rule. A glossary in a .md file keeps it.

This is not a fringe practice. AGENTS.md "is a simple, open format for guiding coding agents, used by over 60k open-source projects."⁸ A plain-text context file the tool reads is already a standard convention — and a glossary is the same shape, scoped to vocabulary. It is a paragraph of agreed meanings you hand to the model before it speaks your language back to you. If you already keep your reusable prompts as plain-text files you own, the glossary sits beside them.

Frequently asked questions

How do I get an AI to use my company's terminology instead of generic words?

Write a glossary file and paste it in as context: each term, your one-line definition, and what not to confuse it with. The model reads your definitions and uses them for that conversation. No fine-tune required, because vocabulary is exactly the kind of context a model can consult during generation to improve the relevance of its output.²

Do I need to fine-tune a model to teach it my jargon?

Usually no. Fine-tuning changes a model's behavior or skill at scale and costs training data, compute, and expertise. Teaching it your wording is a context job, not a training job. A glossary in the prompt handles terminology drift for a fraction of the effort; reserve fine-tuning for permanent behavior change you genuinely need.

What is the difference between a glossary, RAG, and fine-tuning for domain terms?

A glossary steers wording, the lightest fix, a file you paste in. Retrieval (RAG) grounds facts by pulling in your documents at inference time and citing them.³ Fine-tuning changes behavior or style through training. They solve different problems: vocabulary, truth, and skill, in rising order of cost.

How long can my glossary be before the AI ignores it?

Keep it short, ideally one page. The glossary shares the context window, "the amount of text, in tokens, that the model can consider or 'remember' at once,"⁵ with the rest of your prompt. A bloated term sheet gets crowded out. List only the terms that actually drift, and drop anything the model already gets right.

Will the AI remember my glossary between chats?

Usually not. A pasted glossary lasts the session and is gone in the next one. To make it persist, keep it where the assistant reads it every time, inside a custom-instructions or AGENTS-style file the tool loads automatically, and re-attach it when you start fresh. Paste once is not taught forever.

Does a glossary make the AI more accurate about my data?

No, it steers wording, not facts. A model can give a fluent, confident, wrong answer using all your correct terms, because hallucination is content that is "fluent and syntactically correct but factually inaccurate or unsupported by external evidence."⁴ For factual accuracy you need retrieval that grounds answers in your documents; the glossary only fixes the words.

A glossary shapes the words going in. If you also want to catch the answers coming out, that is the job of a separate file — a plain-text eval you re-run to check that a prompt still works. The glossary fixes vocabulary before the model speaks; the eval verifies behavior after. Both are files you own, and neither asks for a pipeline.

A glossary will not make a model truthful, and it will not survive a session it was never re-supplied to. What it will do is stop you from correcting an assistant on your own words for the hundredth time — and because it is a file you own, it does that everywhere, in every tool, long after the one you wrote it in is gone.

This builds on the instinct Steph Ango named as file over app: the meaning of your work should live in a file you control. To keep your glossary in plain text, on your own device, where it travels to whatever assistant you use next, write it in mnmnote.com.