# Reference Breakdown Read this when a supplied video, or a link the user asked to work from, is evidence for what the new video should preserve, adapt or learn from. The useful result is a timecoded reading: what the piece communicates and how its concrete choices make that work. Keep observed facts apart from interpretation, and never claim to have watched footage you only have notes about. ## Get the file ```bash "${PVX}" media fetch "https://example.com/watch?v=ID" --to projects//reference/source.mp4 "${PVX}" media probe projects//reference/source.mp4 ``` `fetch` needs an installed `yt-dlp` (`./local-tools.md`); it never installs anything and refuses to overwrite. Respect the source's rights: a reference is studied and adapted, not republished. Record the URL, duration, aspect and audio presence in the project notes. ## Read the whole through its details ```bash "${PVX}" media transcribe projects//reference/source.mp4 --language en --to projects//reference/words.json "${PVX}" media tile projects//reference/source.mp4 --every 1 --columns 6 --cell 240 \ --words projects//reference/words.json --to projects//reference/overview.jpg "${PVX}" media boundaries projects//reference/source.mp4 --threshold 0.3 "${PVX}" media tile projects//reference/source.mp4 --start 6.8 --end 8.4 --every 0.1 --columns 4 \ --cell 320 --words projects//reference/words.json --to projects//reference/list-change.jpg "${PVX}" media frames projects//reference/source.mp4 --at 6.9,7.3,7.8 --label-time --to projects//reference/frames "${PVX}" media cut projects//reference/source.mp4 --start 6.8 --end 8.4 --to projects//reference/list-change.mp4 ``` Widen the range to follow an argument or a persistent system; narrow it to trace an entrance, a change or a handoff; enlarge a frame to read typography. Word labels on the grid connect picture changes to the words they answer. `boundaries` gives mechanical change candidates, never editorial cut labels; confirm each with frames. ## Write the notes Keep two files under `projects//reference/`: - `ANALYSIS.md`: what the piece is trying to do, how the hook, argument and payoff work, which visual/sound systems recur (host, captions, board, B-roll, stickers, music), and what makes it hold attention. Facts and interpretation stay distinguishable. - `TIMELINE.md`: sections named by source seconds and a phase, each connecting the active objects to the words or actions they serve, with entry, state, persistence and exit. Cite the evidence file for each claim. Example: ```md ## 6.07–8.27 · The workload accelerates The spoken list "videos / voiceovers / ads / scripts" brings one full-frame illustration and a marker-style word per noun, each replacing the last. Labels pop to size, settle, and leave with their picture on the next word. Evidence: list-change.jpg, list-change.mp4; words in words.json. ``` ## Turn the reading into direction Record the relationship, not the seconds: "the board icon lands when the verdict word is spoken", "B-roll covers the explanation and leaves before the next claim". In the new script those become selections and moments (`./semantic-script.md`); the new performance sets the seconds. Then decide, per system, whether it is preserved, adapted or replaced when the person, product, wording, language or platform changes. Transform a role everywhere it appears: a presenter is also an avatar, a voice, a lifestyle B-roll subject and the identity inside a sticker. A product is a prop, a screen, a claim and the closing call to action.