Input: <item_dir> with talks.json + transcript/transcript.tsv. See video-digest for layout/config.
1. Deterministic prep (scripts/)
python3 scripts/vd_sessions.py <item_dir> [--clips] # slices + manifest.json (+ per-talk clips)
python3 scripts/vd_clean.py <item_dir> # cleaned transcript, glossary from digest.json
vd_sessions.py: per talksessions/NN_slug.txt("[HH:MM:SS] text" lines, video offsets) +manifest.json.--clipsalso cuts lossless per-talk mp4s (only useful for multi-talk streams; skipped automatically when a talk spans the whole video).vd_clean.py: collapses Whisper loop/applause runs, appliesglossary/nonspeechfrom the project'sdigest.json(walks up from item_dir) →transcript/transcript.clean.txt,sessions/*.clean.txt,transcript/corrections.md.
2. LLM steps (prompts in references/; fan out ~5 talks per agent)
| Prompt | Reads | Writes |
|---|---|---|
summarize-prompt.md |
sessions/<name>.txt |
sessions/<name>.md (Summary, timestamped Highlights, quotes) |
takeaway-prompt.md |
sessions/<name>.md |
sessions/<name>.take.json {"tldr","takeaway"} |
translate-prompt.md |
highlights.json + take.json | zh.json — only if project languages includes "zh"; run AFTER the slides stage (needs highlights.json) |
Fill {ITEM_DIR} and {AUDIENCE} (from digest.json audience) in prompts.
Order note
summarize → takeaway can run right after this stage's scripts; the slides stage then extracts frames at each highlight timestamp. Translation last (it reads highlights.json from the slides stage).
Glossary loop
After the pilot item, read transcript/corrections.md + skim a sessions/*.md for recurring ASR errors (product names, speaker names) → add regex pairs to digest.json.glossary, re-run vd_clean.py. Only unambiguous fixes; never risky common-word swaps.