The glossary is a side effect. The conflicts are the finding.
Building a glossary by hand is tedious enough that it usually gets skipped until a dispute forces it. By then, Finance has been calling something a "customer" that Sales calls an "account," for months, and nobody noticed until a requirement built on one definition broke against the other.
What it does
- Scans systematically. Capitalised multi-word terms, acronyms, quoted terms, inline definitions, system names, and status/state values.
- Keeps the source's own wording. Definitions stay in the business's language; where a meaning is only implied, it's marked
[INFERRED]. - Surfaces conflicts, never merges them. The same term defined two ways is the highest-value finding — it names both definitions, where each came from, and who should arbitrate.
- Flags undefined terms in use. Words used as if everyone agrees what they mean, but never actually explained — these become your next elicitation questions.
- Sourced and alphabetised. Every entry cites the document and section it came from.
Example
For illustration only, no real client documents involved.
Doc A (Sales): "An Account is any company we have an active contract with." Doc B (Finance): "A Customer is any entity that has ever been invoiced, active or not."
CONFLICTING DEFINITIONS Term: "Account" / "Customer" (used interchangeably elsewhere in notes) - Sales (Doc A): active contract only - Finance (Doc B): ever invoiced, active or lapsed Who should arbitrate: whoever owns the customer data model. Suggested canonical wording: [PROPOSAL — needs sign-off]
How it works with Claude
Install once in Claude — around 10 minutes.
Paste in the source documents you want scanned.
Get the glossary back, with any conflicts called out separately.
FAQs
Why does a project glossary matter?
Because the same term often means different things to different stakeholders on the same project, and that mismatch usually surfaces as a requirements dispute much later than it should.
Can AI catch conflicting definitions across documents?
Yes, when it's instructed to flag conflicts explicitly rather than quietly picking one definition or blending two into a compromise wording that neither source actually said.