certifiedwebmaster/roofgrid/rough-cut

Rough cut

Stage 2 of the 7-stage pipeline (see project CLAUDE.md). Input: one raw video file. Output: projects/<slug>/roughcut.mp4 + transcript.json + cutlist.json.

Steps

1. Pick a project slug from the footage (kebab-case, e.g. youtube-intro).
2. Pick an editing style before running anything. Different content

wants a different cut rhythm, not just different downstream graphics — check /format-presets and pass --preset <name> (see presets/*.json's pacing block). If nothing fits, ask the user rather than defaulting silently; a punchy YouTube-intro threshold will chew up natural thinking-pauses in a teaching/course video, and vice versa.

1. Run:

``bash python3 scripts/rough_cut.py "<path to raw footage>" --project <slug> --preset <preset-name> `` This transcribes with whisper (word-level timestamps), then cuts every silence gap longer than the preset's threshold and every filler word (per the preset's filler-mode) in a single ffmpeg pass so audio and video never drift out of sync.

1. Tell the user where roughcut.mp4 landed and roughly how much runtime

it dropped (cutlist.json has original_duration_sec / cut_duration_sec).

1. Do not proceed to the graphics stage automatically. The user reviews

the roughcut first — this mirrors the no-auto-commit spirit of the rest of this project. If they ask for adjustments ("that cut is too tight on word X", "leave in that pause", "this feels too choppy" / "too slow"), that's a pacing problem — re-run with different --silence-threshold / --pad / --filler-mode (or add a new named preset if it's a recurring style, see /format-presets), don't re-transcribe from scratch unless the source audio changed.

Flags worth knowing

• --preset <name> — loads presets/<name>.json's pacing block as the

defaults below. Explicit flags still win over the preset.

• --silence-threshold <sec> (default 0.6, no preset) — shorter cuts more

aggressively.

• --pad <sec> (default 0.08, no preset) — breathing room kept around each

kept word.

• --filler-mode minimal|aggressive (default aggressive, no preset) —

minimal only strips um/uh/erm/hmm; aggressive also strips like/you know/i mean/sort of/kind of. Aggressive is right for punchy short-form; minimal is right for content where those words are often real phrasing, not disfluency (teaching/course content, interviews).

• --extra-filler "so anyway" (repeatable) — add a filler phrase beyond

whichever set --filler-mode selects.

• --model tiny|base|small|medium|large (default small) — bigger = slower but

more accurate word timestamps; bump up for noisy audio or heavy accents.

After this stage

transcript.json (re-timed to the CUT video, not the original) is the input for every downstream stage — the graphics pass (talking-head-recut), the captions pass (embedded-captions), and the format presets in presets/. Never re-transcribe the roughcut from scratch; those skills should be pointed at this transcript so word timings stay consistent across the whole pipeline.