Monday morning, this is in your inbox.
From: Clinic Director
Subject: visit dataBefore the board meeting next month, can you pull together something on our visit patterns from the last two years? I want to understand what's really going on with our patients. Whatever you think is most useful. Thanks!
It is friendly, it is urgent, and it contains no question. "Visit patterns" could mean who comes in, or how often, or for what, or when. "Really going on" could mean a money problem, a staffing problem, or plain curiosity. The audience might be the director, the board, or somebody the board answers to. Every one of those readings leads to a different month of work, and the email supports all of them.
Answering the wrong reading well is the most expensive mistake in data work, because everything downstream inherits it, and it is the mistake an AI cannot catch for you. The AI answers whatever question you settle on. Settling on the right one happens here, before any file is opened, and it mostly happens by asking.
You get the director for three questions before she disappears into meetings. Choose them well, then choose the brief her answers support.
Check yourself
Think it through before you open the answers.
The question that unlocked the most was about the decision, what will change based on this work. Why does that one question do so much?
Because a decision gives every later choice a test. Knowing the work feeds a care-coordination hire tells you repeat visitors matter, tells you which columns count, tells you what the four summary sentences should be about, and tells you what to leave out. Without a decision, every cut of the data is equally defensible and none of it is aimed. The request said "whatever you think is most useful." Useful has no meaning until you know what it is for.
The director answered the meaning question with a shrug. You tell me what the data can say. What did that answer actually give you?
Permission, and a warning. Permission to propose the framing yourself, which is a normal part of the job and much better discovered now than after a month of work. And a warning that nobody upstream has thought this through, so whatever you propose becomes the plan. When the framing is yours, writing it down and getting a yes on it before doing the work is what protects both of you. That written framing is exactly the brief you produce below.
The job. Turn the director's email into a brief you could hand back to her, one page at most, and specific enough that she could reply "yes, that" or "no, more like this." Getting that yes before opening any file is the whole move.
What the brief needs:
- The question. The director's request, restated as something with an answer. "Which patients account for most of our visits, and how are we serving uninsured patients" has an answer. "Understand what's going on" does not.
- The audience and the decision. Who reads this, and what they will do with it.
- What the data can and cannot answer. These are visit records. Name two questions records like that can answer, and one thing the director might want that they cannot.
- What you still need to ask. Working through the brief probably surfaced more questions. Write the two you would ask the director next.
Compare their approach to yours. How did they take the request apart, and where does that differ from how you did? Look for the move they made that you didn't, and the move you made that they skipped. One of those differences usually matters, and the useful sentence to write is which one, and what picking it up would have changed about your brief.
Compare their brief against yours. Same email, two briefs. Do they aim at the same question? Does theirs promise anything yours holds back, or hold back anything yours promises? Where the two disagree, decide which version the director is better served by, because that, and only that, is the test a brief answers to.
The framing decides the work. An AI will do whatever you ask with equal confidence, which is exactly why the framing is the part that stays yours.