Your first lesson
Play one lesson from start to finish while staying the author of your answers. Allow 15 to 20
minutes once Studium is installed. The lesson shipped with the
repository, atomistique-l1/lesson.json (first-year atomic structure), is enough; no API key is
needed. It is written in French: the screens and feedback follow the lesson's language. Terms such as
scene, hint or harness are defined in the glossary.
Route A: the VS Code lesson pane
- Install the extension as described in VS Code.
- Run Studium: Open lesson, or Studium: Pick a course to choose among the lessons found.
- Read the introduction, open the assets, then continue to the first question.
- Answer yourself; ask for a hint if useful. At a checkpoint, declare your own confidence.
- Read the feedback, shown as its own step, before moving on. After an open answer that no judge assessed, rate your recall with the buttons (again, hard, good, easy): this sets your next review and never changes your mastery.
- At the end, the completion screen lists what you consolidated and when your next review falls.
The notebook is another surface: Studium: Open lesson as notebook plays the same lesson cell by
cell. The pane and the chat participant (@studium in the chat view) are described in
VS Code.
Route B: an assistant over MCP
Connect your harness as in Connect your harness, with BUNDLES_ROOT
pointing at the repository's fixtures/ folder, as the example configurations in
docs/examples/ do. Then send:
Read
skills/studium-harness/SKILL.md. This is a student session, not a simulation of answers. Startatomistique-l1/lesson.jsonwith Studium. Show each screen as Studium wrote it, one at a time, wait for my answer and send my own words. Never invent an answer, a confidence level or a run result, and do not read the answer keys.
The assistant makes these calls; the identifiers below are templates. start_lesson_session
receives:
{"bundle_ref": "atomistique-l1/lesson.json", "learner_id": "demo-local"}
The reply carries a session_id and a presentation whose display_markdown is the screen to show
as is. Your answer goes back in your words:
{"session_id": "ID_RETURNED_BY_THE_SERVER", "input": {"kind": "reply", "text": "b"}}
{kind: "reply", text} accepts what you would type in a chat: a choice letter, a confidence from 0
to 4, a sentence, or "continue" on a reading screen. After the answer, the reply shows
verdict.display_markdown first, then the next screen. Other useful calls:
| Call | Use |
|---|---|
request_hint |
A hint on the current question, without moving on |
explain_last_adaptation |
Why Studium chose the next scene |
get_operator_context |
The public scene, its assets and the expected answer shape |
end_session |
Stop and keep your progress |
tools/list in your harness is the authoritative inventory of tools; the MCP tools
reference describes each one.
A short judgment exercise
Before changing a parameter of an asset, write a one-sentence prediction. After observing, say what confirms or contradicts it. Ask the assistant for a new case that could break the explanation, within the scope the teacher set. Your final answer stays your own; the viewer and the assistant's text are material to examine.
Check that the session works
- A real call returned the scene and the expected tools.
- An answer of your own produced feedback or a transition.
- The feedback comes from the session you used, without mixing pane and chat.
- Any run is reported as executed, failed or not run.
This checks the connection and the flow, not the effectiveness of the course.
Next
- Learn a topic in a conversation: build a lesson from your own sources.
- A scientific experiment with your agent.
- Tools missing or a scene that does not move: Troubleshooting.