Skip to main content

Build and review

Commands​

./lecture list --since 7d                 # sessions found in the captures
./lecture build last                      # the most recent one
./lecture build 20260110-1000             # by id
./lecture build --from 10:00 --to 11:30   # one session for a time window
./lecture build last --no-vision          # skip slide descriptions (fast)
./lecture doctor                          # database, frames, vision model, vault

How a build works​

  1. Transcript. Speech-to-text rows for the session's time span, from the audio device with the most text. Deterministic.
  2. Keyframes. Frames the recorder saved on a visual change, filtered to distinct slides. Two frames are the same slide only if they look alike and their on-screen text overlaps; slides from one deck share a template, so image similarity alone merges them.
  3. Slide descriptions. One vision-model call per slide, cached, so a rebuild is cheap.
  4. Notes. If the transcript fits one model context, notes are written from it directly. If not, each part is digested first and notes are written from the digests, so a long class is covered to the end.
  5. Cards. From each part of the transcript and from the slide text.

Prompt rules that came from real failures​

  • Two different values for one fact: report both and who gave each.
  • A question that was deferred: say it was not answered. Never infer the answer.
  • A statistic on a slide with no source: write "the slide claims".
  • Names: use the spelling on the slide, not the one from speech-to-text.

Review checklist​

  • Slide count is plausible.
  • Names match the slides.
  • Nothing in the notes answers what the class left open.
  • Each card has one idea and a short answer.