Making a decade of program data answer one question
The situation
Every program lead has the same pile: enrollment by term, section fill rates, success and withdrawal numbers, a folder of schedules. Most have tried to build the spreadsheet that would finally make it mean something, and most abandoned it, because keeping it current cost more than it returned. This workflow does not build that model. It makes an un-maintained pile answer a specific question you are facing now.
Steps
-
Name the decision before you open a single file
Paper, before any tool
Write the question in one sentence, and name the decision it feeds: how many sections of the gateway course to schedule for fall, whether to keep the 7 a.m. offering, which course to put a tutor in. Data without a pending decision produces a pleasant report nobody acts on.
What you only learn by doing it: If you cannot name the decision, stop. This is the step people skip and it is the one that determines whether the output gets used.
-
Export, never retype
Your SIS, scheduling system, or IR dashboards
Pull CSV or Excel exports straight from the systems of record. Retyping introduces errors you will never find, and describing your data in prose invites the model to invent the parts you left out.
What you only learn by doing it: Export more columns than you think you need. Trimming later is free; a second trip to the reporting office is not.
-
Put it in one long table, one row per term
One row per term, one column per thing you track, consistent names across years. Ask for help reshaping it, then look at the result yourself. Messy real data is where a model quietly drops rows it could not parse.
What you only learn by doing it: Have it report the row count and the terms covered after every reshape, and check that number against what you sent. A silent row drop is the most common failure in this whole workflow.
-
Write five standing questions and ask them of every year
OpenAI for Education Anthropic-Claude for Education
Five questions, fixed, asked identically of each term: how full did sections run, where did withdrawals cluster, what changed in the schedule, what happened to the following term’s enrolment, what was unusual. Same questions, every year, no exceptions.
What you only learn by doing it: This is the move the whole workflow rests on. As Petroski put it after doing the same thing to 75 pasta shapes: “When you ask the same questions of everything, you begin to reveal the differences between them.”
-
Compare against analogues, not averages
Averages flatten exactly the information you need. Ask instead which prior terms this one most resembles and on what grounds, then go read what actually happened in those terms — including what you did about it.
What you only learn by doing it: Make it name the two or three closest years and say why. If the reasons are vague, the resemblance is not real and you should not act on it.
-
Write down the call, the date you check it, and what would prove you wrong
A plain text file beside the data
Three lines in a file: what you decided, when you will look again, and the observation that would mean you got it wrong. This is what turns an interesting afternoon into something that compounds next year.
What you only learn by doing it: The third line is the one that makes you better at this. Without a written falsifier, every outcome gets remembered as roughly what you expected.
Where this breaks down
Ten or twenty rows of annual data support a sentence like “this term most resembles fall 2019.” They do not support statistics. If the model offers you a correlation coefficient, a trend line or a p-value on twenty rows, it is performing confidence it has not earned. Ask for comparisons and resemblances, not significance.
It will also find patterns that are not there. Any model asked to find structure in a small table will deliver some, fluently. The guard is step four: you decide the questions in advance, so you are testing a few stable things rather than fishing.
The maintenance trap is the real killer and the reason most attempts die. Whatever you build here has to survive a week when you are buried. If reproducing it next term takes more than an hour, it will not happen.
Student records. Do not paste student-identifiable data into a consumer account. Aggregate first — counts by term and section, never names, IDs or rows per student. If you cannot aggregate it, this is a conversation with your institutional research office, not with a chatbot.
Provenance: adapted from Dan Petroski, winemaker and founder of Massican, interviewed in OpenAI’s ChatGPT for Pros subscriber newsletter, 1 October 2026. Petroski had abandoned a spreadsheet model of grape ripening because the upkeep defeated him, then used 17 years of accumulated vineyard records to characterise an incoming vintage against its closest prior analogues. 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.