LLMs working inside a reasonably large organization run into a wall of specialized vocabulary: terms and meanings that never appeared in the training set, which makes the model less effective at domain tasks. Generating a glossary of that terminology with an LLM and making it available to other LLMs is one way to close the gap.
Mining the organization's own text
Almost any organizational material can serve as source: documents, emails, chat logs. Transcripts of audio or video can help in some cases, though most organizations already produce enough text. From this corpus, the generating LLM should extract both single words and multi-word phrases that habitually appear together as meaningful terms. It should be prompted to flag two categories: terms that are new to it, and familiar terms used in unfamiliar ways — the second kind often produces the most confusion.
Drafting the definitions
With the term list in hand, the agent can write a glossary with a definition per entry. Where a term carries several meanings, it should supply alternative definitions and note the context in which each applies. Asking the LLM to follow a dictionary output format, geared to precision and conciseness, gives a usable draft.
Domain experts must then review the draft and correct the places where the model has misunderstood. The corrected glossary is placed where any other LLM can consult it — in a form that is easy to look up and that minimizes the tokens spent finding a meaning. Two operational habits matter: a schedule for regular review and refresh by both agents and humans, and version control so the glossary's evolution stays visible.
Bounded contexts and multiple glossaries
Large organizations often define the same term in subtly different ways in different places, which is the Domain Driven Design notion of a Bounded Context. Source code is a particularly interesting bounded context, since a code-derived glossary shows how terminology is actually used inside a codebase.
Sometimes one glossary can span several bounded contexts; other times separate glossaries per context work better, letting an LLM consult only what is relevant to its task. A single standard glossary is worth resisting. The value of a glossary is frequently not a definitive definition of a term but the visibility it gives to differing usages across contexts. Where multiple glossaries exist, comparing them — or having LLMs compare them and describe the differences — is instructive, and comparing a source-code-derived glossary against one built from human text is especially so.
For a large glossary, the initial context can carry only the terms and links to their definitions; the model follows the links for the full entries it actually needs.
Using glossaries in agentic engineering
Glossaries feed directly into agent workflows. Code generation can be informed by them so that names in the generated software match enterprise vocabulary, and sensors in a harness can check that usage both inside the software's internals and in its user interface.
Generating code or documents may also create new bounded contexts. Internal names used within an organization, for example, should not surface in a customer-facing UI, where they would most likely only confuse the customer — the customer sits in a different bounded context from the part of the enterprise serving them.
It also pays for a glossary to list terms that should not be used, together with the term to use instead. Deprecated terms can be handled the same way, with entries stating the period over which the term applies.
Where this fits
Conversations with people running agents in large organizations return repeatedly to the difficulty of specialized language. That pressure motivates domain modeling, ontology building, and explicit context layers in enterprise data; LLM-generated glossaries are a simple but useful part of that work.



