The session ask · Artifact
Make one image-generation artifact
Make something useful for your students or your own learning: a tool, a webpage, an activity, or even a well-formed set of questions — a rough list of questions is a real artifact. One route: take an existing tool from the GitHub repo, feed it into an LLM, and ask for the version your classroom actually needs (with credit; pull requests welcome). While you're in the tools, note what works, what doesn't, and what's missing — that feedback shapes the next revision.
~45–90 min · any form counts, however rough
Default · Async share
Run the default test
Prompt an image model with "a doctor" — or your own deliberately unspecified prompt — and read off what the model decided for you: gender, age, race, setting, camera angle. Change one detail and run it again. Log what the defaults were, what your one change shifted, and what stayed stubborn. No image model access? Practice the method on the authored simulations instead.
~15–20 min · one logged comparison and two sentences
Go deeper · Be the model
Run your own Human Diffusion Canvas
Write a three- or four-phrase prompt of your own and run the canvas with someone else — family, friends, colleagues — revealing it phrase by phrase while they draw. Export the GIFs. Then run a second round giving the full prompt up front, and compare: did the drawings still converge? Which shared visual conventions appeared, and where might those defaults come from?
~25–35 min · two or more exported drawings and one observation
Go deeper · Probe
Find where spatial reasoning breaks
Fluent output can hide reliability limits; layout requests make them easier to inspect. Ask an image model for a one-page zine layout (eight panels, top row upside down), or ask a coding model for a pure-CSS image. Compare against an easy prompt like "a red panda." Document exactly where coherence falls apart, what the failure suggests about spatial reliability, and what another test could reveal about whether the problem lies in planning, representation, or rendering.
~20–30 min · one easy/hard comparison, logged
Teach
Adapt the marker version for your room
The canvas activity started on paper: everyone grays their page, each revealed phrase means a new color added on top, then the room compares results. Design the version for your students — what's the prompt, how do you time the reveals, and what's the debrief question about where the shared defaults come from? Name the caveat too: a real model gets all the words at once.
~45–60 min · one-page protocol, slides, or rough worksheet
No AI · Critique
Analyze a frozen example
Choose one authored example from the Image Prompt Pack. Identify one default, assumption, or limitation you can support with specific visual evidence — what's in the frame, who's in it, what the composition assumes. Separate what the image shows from your hypothesis about why, then name the real-world test you'd need to check that hypothesis.
~20–30 min · one annotated example · no live AI use required