The Shape of a Good Question
By Michael Polzin, Regenerative Architect. Published 2026-07-01.
A room learns in the shape of the questions it is asked. Change the shape, and the room changes with it.
Most bad classroom moments do not begin with a bad answer. They begin with a question that had one right answer to begin with, delivered in a tone that made the room understand it was being tested rather than invited. This piece is a small field guide to the opposite: questions that ask a person to show their working model of the world.
What a good question is doing, underneath
In the active-inference frame this site works from (Class E, standard formulation in Parr, Pezzulo, and Friston, 2022), a mind is a model that predicts and gets updated by surprise. A good question is a device for making a person's current model visible to themselves and then to the room. It does not primarily ask, "do you know the fact." It asks, "what shape is your map of this territory right now, and where does it feel uncertain."
Our own facilitator prompts and intake forms are configured to log answers as model snapshots rather than as scored responses (Class C, this is how the /api/intake pipeline and the workshop protocol are wired). When a person's answer is treated as a picture of their current model, a wrong prediction becomes useful data, not a personal verdict.
Four properties of a good question
Open. More than one honest answer will fit. If just one completion is correct, and it is a single word, you have written a cloze. Cloze items have their place. They are not the tool for surfacing a model.
Non-leading. The wording does not smuggle in the expected answer. "Don't you think it would be safer if..." is a request for agreement, not a question.
Model-inviting. The question asks for a sketch of how the person thinks the thing works, not for the label of the thing. "Why does the ice melt faster on the metal tray than on the plate" invites a model. "What is conduction" invites a definition.
Safe to be wrong in. The room has been set up so that a partial or mistaken model can be spoken out loud without a social cost. Without this, every other property collapses; people will simply guess what you want to hear.
Concrete rewrites
The rewrites below are the same content, aimed at the same standard, in a different shape.
Before. "What is the answer to number four?"
After. "Walk me through how you got started on number four. Where did it feel solid, and where did it feel shaky?"
Before. "Does everyone understand?"
After. "If you had to explain this to someone who missed today, what is the one sentence you would lead with, and what part would you leave for later?"
Before. "Who can tell me the definition of a mammal?"
After. "If we found a new animal in a cave tomorrow, what would you look for to decide whether to call it a mammal? What would make you unsure?"
Before. "Any questions?"
After. "What is the part of this you would want to test before you trusted it in front of your own class?"
Notice what the After versions share. They assume the person has a model, give them a way to show its edges, and treat uncertainty as information a facilitator wants rather than a weakness a learner needs to hide.
Three failure modes to watch for
The compound question. Two or three questions stapled together with "and" tell the room that whichever one is easiest to answer is the one they should answer. Ask one.
The performance question. "Who can tell me..." is aimed at the person who already knows and wants to say it. The other people learn to stay quiet. Aim your questions at a model, not at a volunteer.
The reassurance question. "This makes sense, right?" is not a question. It is a request that the room protect the facilitator's confidence. Ask instead for the specific place the sense breaks down.
Whether these three rewrites, applied consistently over a semester, measurably shift how learners speak about their own uncertainty is (Class U) not yet demonstrated at the scale we would need to make a general claim. We are gathering the data through the workshop's repair-loop logs and will publish the falsifier alongside any positive result (Class F).
A quiet close
You will know the shape has changed when the room starts answering with sentences that begin, "I think it works like this, but I am not sure about..." That sentence is the entire discipline. It is a person handing you their current map and pointing at the fog. Every good next move a facilitator can make starts from there.
Read next
- Human-in-the-Loop, What It Actually Means When a Learner Is in the Room. The cornerstone piece this post sits inside: the learner is the model, and questions are the instrument for making that model visible.
- The Map and the Territory in a Lesson Plan. Where a good question fits in the arc of a session, and how to sequence them so the room's map keeps getting sharper.
- The UNI workshop. The paid, deep-dive program where facilitators build this practice with us, on the record, with the receipts published as we go.