Make a glossary from documents without inventing definitions

Example result.

1. Choose the documents and the reader

Pick a small, coherent set of documents and decide who the glossary is for. A new colleague needs different terms from an experienced specialist, and extracting every capitalised word produces a long list that explains very little.

Check the version and status of each document. Don’t mix an old draft’s definition with the current one without showing the difference. Keep document names and section headings with the text you paste or upload, so a reader can find the original passage.

Why this works: a named reader tells the model which terms are worth explaining, and document names make every row traceable.

What you should see: two or three documents, each with a name, and one sentence describing the reader.

2. Ask for definitions with evidence

Paste or upload the documents, then send this prompt.

Make a concise glossary for [reader] using only the documents below.

For each useful term, give its meaning in this source context, the document and supporting passage, and whether the meaning is explicitly defined or only suggested by context. Keep different meanings separate. If a term or acronym is used but not defined, write "not defined in these sources" rather than expanding it from general knowledge.

Don't invent document names or merge conflicting meanings. Finish with a separate list of terms the document owner needs to confirm.

Documents:
[document names and text]

[reader] is the person who will use the glossary, such as “a new starter in the warehouse team”. [document names and text] is the material from step 1.

Why this works: the passage field makes each definition checkable, context stops a familiar word inheriting the wrong meaning, and the undefined-term rule keeps common knowledge from posing as the document’s own definition.

What you should see: a table with one row per meaning, a passage for each, and a short confirmation list at the end.

3. Check each row against its passage

Open the supporting passage for each row. Does it define the term, merely use it, or describe one example? Check for a scope or condition the draft explanation may have dropped.

If a meaning is a helpful interpretation rather than a statement in the source, label it “proposed explanation — confirm”. Don’t put it next to an exact citation and let the reader assume the source says those words. Check conflicting uses separately: two documents can legitimately use the same word for different things, and one blended definition misleads both audiences.

Why this works: the model is good at finding candidate terms and weaker at telling a definition from a passing use. Reading the passage is the only way to tell.

What you should see: every row either matches its passage or is relabelled, and no row merges two documents’ meanings.

4. Ask the owner to confirm the open terms

Turn the confirmation list into one short message for the person who owns the documents.

Draft a short message asking [owner] to confirm what these terms mean in their documents: [list of open terms]. For each term, quote the sentence where it appears and name the document. Don't suggest an expansion or a likely meaning.

[owner] is the person or team responsible for the documents. [list of open terms] is the confirmation list from step 2, after your check.

Why this works: quoting the sentence gives the owner the context they need to answer quickly, and leaving out guesses stops the question from steering the answer.

What you should see: a message of a few lines that you can send through your usual channel, with one quoted sentence per term.

Check the result

• Every defined meaning has a document name and a passage you’ve read.
• A word used two ways in two documents has two rows.
• Acronyms the documents don’t expand say “not defined”, not a familiar expansion.
• Qualifications survive: “packed and checked” hasn’t become “released”.

The usual failure: the glossary fills a gap from general knowledge. An undefined acronym arrives fully expanded, or two meanings are blended into one tidy row. Spot it by looking for a row whose passage only mentions the term without explaining it.

These sources use [term] but don't define or expand it. Remove the definition you added, mark it "not defined in these sources" and add it to the confirmation list. Keep the source reference. Where a term has different meanings in different documents, keep separate rows for each meaning.

A second assistant supplying the same familiar expansion isn’t source confirmation. The owner is.

Fictional example. A new coordinator works across a warehouse team and a helpdesk, and asks for a glossary of both teams’ notes.

Warehouse note
Ready means the items are packed and checked. Ready items are not released automatically.
Dock means the loading area assigned to the shipment.
SLA is mentioned in the handover checklist but is not defined here.

Helpdesk note
Ready means the request has enough information to investigate. It does not promise a fix.
Dock means a desktop accessory connecting a display and input devices.
QBR is scheduled; this note does not expand the acronym.

The first draft came back with two problems. Weak: “Ready — items or requests prepared for the next stage” and “SLA — service level agreement (warehouse note)”. The first row blends two teams’ meanings, and the second is general knowledge with a citation attached; the note never expands SLA.

After the repair prompt, with [term] set to SLA and QBR, the glossary reads:

• Split meanings: “ready” and “dock” each get two rows, because a warehouse reader and a helpdesk reader would act on them differently.
• The qualification stayed: packed and checked does not mean released. A draft that says “approved for immediate release” fails, however tidy it looks.
• The gaps went to the owner: step 4 turns SLA and QBR into a two-line question to the warehouse and helpdesk leads, quoting where each appears.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Add a document that uses the term a third way

Glossaries grow. Suppose a Finance note arrives that says “Ready means approved for payment.” The right result is a third “Ready — finance” row, not an edit to the existing two.

Add this document to the glossary: [new document name and text]. For any term it uses, add a new row for its meaning in this document. Don't change existing rows unless this document is a newer version of their source; if it is, show the old and new wording side by side.

Then check that the existing rows still match their passages.

Separate a quotation from a plain-language explanation

You may want a short paraphrase for the reader and an exact quote for review. Keep them in two fields. Compare the paraphrase with the full passage so it doesn’t lose an exception or change the meaning. If a term needs specialist interpretation, such as a contractual or clinical meaning, ask the owner or the right professional rather than letting the glossary settle it.

Keep the glossary current

Record the document version beside each row. When a document changes, recheck the affected rows rather than adding a second, unchecked version next to the old one. Keep open terms visible in their own list; don’t delete them to make the glossary look complete. For general AI vocabulary, the site’s AI glossary has a different job, and the output-format guide helps you turn the result into a shared page.

Sources

• Upload files to Claude
• File uploads FAQ (ChatGPT)