Finding the handful of words that define what you do
The situation
A program, a service or a product accumulates outside description: reviews, survey comments, client feedback, what partners say about you. Read one at a time it is noise, and the loudest items distort memory. Read as a corpus it has a signature — the few things people independently keep saying. This finds that signature and turns it into language you can use in a catalog description, a grant narrative or a pitch.
Steps
-
Gather outside voices only
A folder, a document, or an export
Collect what other people wrote: reviews, open-ended survey responses, partner feedback, press. Nothing you or your office authored. Your own marketing language in the corpus guarantees you will rediscover your own marketing language.
What you only learn by doing it: Keep the dates. You will want to split early years from recent ones later, and you cannot reconstruct that afterwards.
-
Keep the corpus honest
Your text editor
Include the lukewarm and the negative. A lexicon built from favourable comments only describes your best day, which is the one thing you did not need to find out.
What you only learn by doing it: Count what you excluded and why, in one line at the top of the file. If the exclusions are hard to justify in a sentence, put them back.
-
Ask for recurring vocabulary, not a summary
Gemini Notebook (formerly NotebookLM)
Ask which words and phrases recur across the whole set and roughly how often, grouped into themes. A summary gives you one smooth paragraph; what you want is the raw repetition underneath it.
What you only learn by doing it: Ask it to quote two real examples for each recurring term. If it cannot produce the quotes, the term came from the model and not from your corpus.
-
Cut to six, then test against something it has not seen
Anthropic-Claude for Education
Force the list down to about six words. Then take a review you held back and check whether those six would have predicted it. If they would not, the list is aspiration rather than signature.
What you only learn by doing it: The cut is the work. Twelve words describe everything and distinguish nothing, which is how you end up with language indistinguishable from every other program’s.
-
Put the six words where they do work
Wherever your real copy lives
Into the catalog description, the program review narrative, the advising one-pager, the grant abstract. A signature that lives in a file you never reopen has produced nothing.
What you only learn by doing it: Note the date you built it and set a reminder to re-run in a year. Quoting a four-year-old lexicon as current description is how a program comes to describe a version of itself that no longer exists.
Where this breaks down
Word frequency measures what people wrote, not what is true. A term recurs because it is salient, or because the prompt invited it, or because one influential reviewer set the vocabulary everyone else borrowed. The output is a description of the discourse, which is useful, and not a measure of quality, which it will resemble.
Survivorship bias is built in: the people who stayed are the ones who commented. A signature drawn only from completers is a signature of a self-selected group, and saying so in the write-up costs you nothing.
Do not run this on student evaluations of a named instructor. Student evaluations of teaching carry documented demographic bias, and aggregating them into a vocabulary about a person launders that bias into something that looks objective. Course-level signal is a fair use; personnel use is not.
Signatures drift. Anything older than about five years is history, not description. Date the corpus and re-run it rather than quoting a lexicon you built once.
Provenance: adapted from Dan Petroski, interviewed in OpenAI’s ChatGPT for Pros newsletter, 1 October 2026, who ran 15 years of third-party reviews of his wines through ChatGPT to find the recurring aroma and flavour vocabulary, producing what he calls a flavour lexicon: “It sums up my whole history of a wine in six words.” The sequence, the checkpoints and the cautions below are our construction, written for a classroom. Treat the source as one practitioner’s documented experience in a vendor publication, not as evidence that this works generally.