Andreas, 2 October 2026: "we now have a spectacular use case. This is very good, and it should be the example of how to set it up." This page is that example. It lists what was done, in order, with the files it made. Every step is a button in the app and a command underneath. The course is courses/micro101, built from the OpenStax book Principles of Microeconomics 3e (CC BY-NC-SA 4.0).
Each step is drafted by the model the teacher chose, Claude Opus here. Its gate checks it. It is left as a draft for the teacher. The session's next-step button always says what comes next.
| Step | What it makes | Gate | Time, cost |
|---|---|---|---|
| Draft the plan | sessions/s01/plan.md: goals, rubric rows, seven segments, the worked example with Alphonso's $10 | citations resolve in the index | about two minutes, under $2 |
| Draft the slides | decks/L1_*.tex, 30 frames, 28 in the video, 109 steps | builds three ways, zero overfull boxes, the scanners | about ten minutes, under $5 |
| The adversary | sessions/s01/adversary.md: a second model reads the transitions and the difficulty | findings with Agree and Dismiss | a few minutes |
| Draft the narration | scripts/s01_*.md: one block per slide, [click] at each step, 2205 words | steps match the slide pages, both prose checks | about five minutes |
| Voice | audio/s01/, 109 clips in the teacher's own voice | the transcript check, 0.953 on average | about 18 cents |
| Video | video/s01_async.mp4, 13 minutes 17 seconds, and its compressed copy | one page per clip | about two minutes |
| Draft the notes | notes/s01.qmd and its PDF, nine pages | review questions at the end, no solutions, no verbatim narration, both prose checks | about five minutes |
| Draft the quiz | quizzes/quizdata_s01.py and the quiz, ten questions | the hint rule, the numbers checked | about three minutes |
| Draft the problem set | decks/A1_.tex and decks/S1_.tex, six tasks | builds, every rubric row trained | about five minutes |
| Draft the in-person deck | decks/IP1_*.tex, 21 frames, nothing the video did not cover | builds, the scanners | about five minutes |
| Write the companion | inperson/s01_companion.tex, 11 A5 pages | builds, page 1, both prose checks on the text read back | about five minutes |
| Speaker notes and PowerPoint | scripts/s01_speaker.md, decks/L1_*.pptx | one note per slide | about three minutes |
| Make the survey | segments/s01_segments.json, sessions/s01/survey.md | options equal the segments | instant, no model |
Times and costs are what the runs of 1 October took on one Mac with Claude Opus. They are not promises.
Read each draft in the app and press Approve, or edit it in the block editor and let it go back to draft. Three things the agents could not decide came back as questions. The letter for income. Whether the production possibilities frontier belongs in session 1. A chapter placed in the in-person lane that the video does not cover. Pronunciation is fixed by listening and choosing. pipeline/pronounce.py made five spellings of "economics" in the teacher's voice, and the chosen one went into lexicon.json.
Then Settings, Publish: a dry run, a tick, and the push to TavlAI or Canvas, where everything arrives unpublished.
sessions/s01/plan.md, decks/L1_async.pdf, scripts/s01_scarcity_choice_and_the_.md, video/compressed/s01_async.mp4, notes/s01_notes.pdf, quizzes/s01_review_quiz.md, decks/A1_.pdf, decks/S1_.pdf, decks/IP1_.pdf, inperson/s01_companion.pdf, decks/L1_.pptx, sessions/s01/survey.md. The PDFs, the audio and the video are built products. They are not in the repository. The sources are.