Why a graph teaches better
A prompt history is a transcript of guesses. It shows what somebody typed, not what they understood, and a student reading it learns to imitate wording rather than to reason about inputs. Worse, it rewards the appearance of fluency: the student who found a lucky phrasing looks more competent than the one who reasoned carefully and got a worse image.
A graph shows the structure of a decision. This reference feeds this shot. These settings produced this result. Changing that node changes these three and not those two. It can be read, critiqued, marked, and — crucially — compared between two students who attempted the same brief.
That also makes the failure legible. When a student's set drifts, you can point at the node they should have shared and did not, which is a teachable observation rather than "try a different prompt".
In a class
Guest sessions for a lab
The studio opens without accounts, so a room of thirty students can start in the first two minutes rather than the first twenty. Guest runs return bundled samples, which is enough to teach structure without spending anything on generation.
Workflows as assignments
Publish a graph as a starting point and have students extend it. What they changed is visible in the structure, which makes marking a matter of reading rather than guessing.
Reproducibility
Settings and inputs stay attached to results, so a claim in a paper or a portfolio can be checked rather than taken on trust.
Critique that transfers
The vocabulary students build here — reference, dependency, shared input — is the vocabulary of every production pipeline they will meet afterwards.
A shape that works
One suggestion, from how the tool is actually built rather than from pedagogy we are not qualified to invent.
Week one — one node
A single generation. Read the settings, change one, run again. The lesson is that a result has causes, not that a prompt has magic words.
Week two — two nodes and an edge
One reference feeding one output. Change the reference. This is the entire idea of the tool, and it is worth a whole session.
Week three — a set
One reference, six outputs. Students discover drift by building it wrong first, which teaches it better than being told.
Week four — a brief
A real constraint with a deliverable set. Mark the structure alongside the images.
Access for cohorts
If you are teaching or researching in this field and need seats for a cohort, write to us. There is a programme for this and it is not a discount code — tell us the course, the number of students, the term dates and whether you need real generation or whether sample mode covers what you are teaching.
For research groups: reproducibility is the part we care about most, and we will help you set up a project structure that a reviewer can actually verify.
Questions from institutions
Do students need accounts?
Not for guest mode, which covers structure teaching entirely. Real generation needs an account, because a cookie anyone can mint is not something that can be metered.
Can we run it on our own infrastructure?
Not yet. On-premise deployment is on the roadmap under Later. If institutional policy requires it, tell us — that is exactly the input that moves something up the list.
Is student work used for training?
No. No projects are used as training data, student or otherwise.
What ages is this appropriate for?
The tool has no age gate, and generative models can produce unexpected output from innocuous prompts. For under-16 teaching we would suggest sample mode and a prepared workflow rather than open generation.
Can we get material to teach from?
The documentation is written to be readable by a student, and you may use it in course material with attribution. If you want something structured as a syllabus, write and tell us what you need.