--- name: rough-cut description: Editorial rules for choosing what to keep in talking-head footage (retakes, false starts, spoken notes to the editor, fillers) and the keep-list / moments formats that a planner script turns into frame-exact cuts. Use when deciding a cut, revising one from feedback, or writing keep.json. --- # Rough cut: what stays You decide **what** stays, by word index. A planner script decides **where** each cut lands (pads, waveform checks, pause shortening, frame snapping). Never do its job by hand. The brief creators give: **remove retakes and silences, keep it feeling like a natural vlog.** Not a jump cut every second. ## Read first The manifest, then every clip's word list in order. Each token: `{i, text, start, end, type}` with type `word` or `audio_event` ("(laughs)", "[lighter click]"). Read it as a story before deciding anything. ## Rules 1. **Retakes:** keep the best complete take, usually the last. Never keep a take that trails off. Cut the restart cue with it ("let me do that again", a swear at themselves). 2. **False starts:** cut an abandoned sentence plus its restart cue. 3. **Spoken notes to the editor** ("editor, roll the video here", "this slide is a video"): cut them and write each as a `note` moment → becomes a pink marker on the join where it was. 4. **Fillers stay.** It's a vlog. Only cut a filler that is clearly dead air on its own. Stumbles inside a word stay (cutting them makes a visible jump). 5. **Keep their voice:** asides, jokes, a laugh, a deliberate pause before a point. 6. **Pauses are not your job.** Give whole thoughts as one range; the planner shortens long pauses. 7. **Keep recorded order** unless asked. 8. **Unsure? Keep it and flag it.** A wrong cut is invisible to the creator; a flagged keep is a 5-second decision. Typical flags: near-identical lines in a row (deliberate repetition or retake?), self-corrections, a joke about "starting this three times". ## keep.json ```json { "segments": [ {"clip": "C0001", "from": 12, "to": 57, "why": "opening, second take"}, {"clip": "C0002", "broll": true, "in": 0.0, "out": 4.0, "why": "camera goes into the bag"} ], "flags": [{"clip": "C0001", "at": 141, "note": "kept a stumble, check it"}], "summary": "3-6 lines in plain words" } ``` `from`/`to` are inclusive word indices. B-roll segments are by seconds (see `broll`). Optional per range: `lead_in` / `tail` seconds of room (see `vlog-pacing`). ## moments.json ```json {"moments": [ {"type": "note", "clip": "C0001", "from": 1964, "to": 1994, "label": "VIDEO: product demo", "note": "Roll the demo here. They said: '…'"}, {"type": "punch", "clip": "C0001", "from": 812, "to": 818, "label": "the line that lands"} ]} ``` - `note`: covers the cut words; the marker goes on the join. - `punch`: a real punchline (conclusion, payoff, strong opinion), about one every 2–3 minutes. The picture pushes in to 115% **for that line only**, then goes back. ## After planning Read the review file once: struck-through text must be only what you meant to cut. Report: raw vs cut length, pieces, every retake removed (timestamp + gloss), every note, every flag. ## Revisions Creators speak in timeline times ("cut 02:14 to 02:20"). Map to word indices with the review file, edit the keep-list, re-plan, rebuild as a new version, and report only what changed.