How a capstone works in Ulern
Most programmes end with a test of what participants remember. A capstone ends with something real: a memo, a plan, a report that someone could act on. That is where a lot of learning is meant to show, and it is also where a supervisor’s time runs out, because every draft needs reading.
This post walks through how a capstone works in Ulern, using the video above. The example is a one-page decision memo: a manager takes a decision from their own work and writes the memo that settles it. Everything on screen is the product, recorded in October 2026 in a demo academy we keep for this. The people are invented, and so are the participants’ drafts (written by simulated participants). The plan, the tutoring session and every review in it are Ulern’s own.
The supervisor sets the method
Before anyone starts, the supervisor decides three things: what the participant will make, the rubric it is judged by, and the method, step by step. Some steps are work the participant writes. Others are gates, where a supervisor has to sign off before the next part opens. The learning plan sits around the project, with milestones that teach the reading the memo needs.
In the example, the memo is judged on four weighted criteria: a decidable question, evidence read fairly, an actionable recommendation, and honesty about uncertainty. The method has five steps: frame the decision, gather and judge the evidence, an evidence check by the supervisor, the recommendation, and a final review. Each step names what it is judged against, so the participant can read the standard before writing a word.
Publishing fixes a version. A participant keeps the version they started on, so nobody’s work moves under them when the supervisor improves the method later. Then the supervisor invites participants: one by one, a whole cohort at once, or by email.
The participant learns, then makes
The participant starts with their plan and sees where they stand on every milestone. Teaching happens in sessions with the AI tutor (on screen it is called Uli), and it works on the participant’s own situation. In the video, an operations manager opens with “I have a routing decision to make at work”, and every question that follows is about choosing and reading the measures for that decision. Each answer is checked before the session moves on, and when he writes, the AI tutor says what holds and what still needs checking.
Then the participant writes each step of the deliverable themselves. When they ask for a review, the AI tutor reads the draft against that step’s criteria, marks each one, and names what is still open. It never writes the draft. A draft the AI tutor wrote would be the AI’s work, and the supervisor would end up grading a model instead of a participant.
A supervisor decides at every gate
At a gate, the work goes to the supervisor, not to the AI. In Studio, the supervisor sees what is waiting on them, opens the participant’s work, and reads the AI tutor’s checks beside the draft. Then they approve the step or return it, and a return always carries a reason. The supervisor’s decision is the one that counts, whatever the AI tutor thought of the draft.
The method caps how many times a step can come back, so a return is a second chance, not an endless loop. The reason lands in a thread beside the work, where the two of them can talk it through. In the video, the supervisor sends back a support lead’s evidence on a four-day-week pilot: the pilot ran in summer, her quietest quarter, so he asks her to set the same months of last year beside it. She asks one question in the thread, he answers, and she rewrites. The AI tutor reads the new draft, and the gate waits for the supervisor again.
What the participant keeps
When the last gate is signed, the participant puts the document together. The AI tutor can smooth the joins between sections, but every change to their own words is shown to them and needs their answer, so nothing is exported unseen. They download the finished memo as Word, PDF or Markdown, and it is theirs to keep.
At the end of the file is the record of how it was made: the steps, the AI tutor’s reviews, the rounds with the supervisor. It goes with the document, not instead of it, and the supervisor sees the same record.
Why it works this way
The scarce resource in a capstone is the supervisor’s attention, not the software. So the AI tutor takes the first read of every draft, against criteria the supervisor wrote, and the supervisor spends their time on the decisions that are theirs to make: whether this step is good enough, and what to say when it is not.
The record keeps the AI’s part visible. Anyone reading a finished memo can see who wrote it, what the AI tutor said about it, and who signed it off.
If you run a programme that ends in real work, book a demo and we will show it on a project like yours.