Skip to main content

​
The short version

An AI model produces text. An agent combines a model with a fixed job, tools, state, and checks. The Question Studio agent has one bounded job: turn the source pages or topics you selected into reviewable multiple-choice drafts for a specific module. Curators retain publishing control.

​
What happens during a run

1

Load the authorized source

Question Studio loads the processed document, destination module, selected pages or topics, and existing module questions. The existing questions support duplicate checks.
2

Plan the work

The system assigns question slots across source-supported topics, objectives, and requested types. Page-based runs may build this plan directly or with a planning model, depending on the mix.
3

Draft in parallel

Small workers receive bounded page excerpts and question contracts. Candidates appear in the interface as workers finish.
4

Check the candidates

LearnTerms checks source evidence, answer structure, and duplication. Some run paths also include an independent model review; other candidates are explicitly labeled as locally checked.
5

Hand control to the curator

You inspect, edit, select, and save candidates. Saved questions remain drafts until someone publishes them from the module.

​
Terms you will see

​
What the agent can read

The agent reads the selected source scope and existing questions needed for duplicate checks. Question Studio runs operate without open-web access. The original PDF page can help you recognize diagrams and layout, while generated questions rely on extracted text and retrieved excerpts. Treat visual facts missing from extraction as unavailable to the agent.

​
Limits of automated checks

Automated checks can catch:
  • missing or weak source evidence
  • ambiguous answer structure
  • repeated or highly similar questions
  • mismatch between a requested question type and the available source
  • malformed question fields
Curator review remains required for educational quality, clinical correctness, and extraction fidelity. A locally checked candidate deserves especially careful review of the correct answer and distractors.

​
Why runs can return fewer questions

Question Studio treats the requested count as a target. A worker can abstain or a candidate can be cut for an unsupported type, thin evidence, an ambiguous answer, or duplication. Fewer defensible drafts are better than filling the batch with unsupported questions.

​
Your role in the loop

You remain responsible for the final editorial decision:
  1. Check the cited evidence.
  2. Verify the correct answer and closest distractors.
  3. Edit wording or rationale when the candidate is fixable.
  4. Save only strong candidates as drafts.
  5. Publish from the module after a final review.