# Learnings from real edits Short notes on what went wrong on real videos and the rule that came out of it. The skills already include these; this file is the "why". ## Cutting - **Cut from words, not from volume.** A volume-only silence cut gave ~800 pieces on a 34-minute class; the transcript cut gave 120 and was loved. - **Classes like it tight, vlogs don't.** 0.5 s max pause / 0.3 s kept was perfect for a class and "popping" for a vlog. Vlog: 1.0 / 0.6. - **Scenes need room.** Starting 0.1 s before the first word and ending right after the last made every scene slam into the next. 0.5 s in, 0.6 s out (more on the ending), stopping before the hand reaches the camera. - **Zoom only on jump cuts.** Zooming on scene changes added pops; a new scene opens at 100%. - **A punch-in lasts one line.** Without a split after the punchline, one push-in held 83 seconds. ## B-roll - **A transcript cut drops silent clips.** Treat 0-word clips (and clips whose "words" are radio lyrics) as B-roll. - **Watch it properly.** 5 frames per clip led to cuts the creator called "weird and wrong". At 2 fps the story appeared: camera into the bag → black → out of the bag at the airport, a free transition. ## Sound - **Music louder where nobody talks.** -19 LUFS on B-roll was too timid; -16 with a slow rise starting in the speaker's tail made the music flow into the B-roll. - **Check the edited music track's ending.** An exact-length edit faded out 25 s early, under the B-roll it was meant to lift. - **Literal foley is cheesy.** Zipper, cloth swooshes and a jet pass-by were "horrible". One soft untuned swell into each B-roll, quiet room tone, and the clip's own sound worked. - **A silent music track can be a template setting.** The template's music track had a track-level AI Music Remixer on; anything on that track was silent. Found only by rendering and correlating, then looking at the track Inspector. - **Normalizing can hurt.** A silent clip was boosted +30 dB into hiss. ## Voice - **"Underwater" voice = measure it.** Channels were identical (not a phase issue); the spectrum had almost nothing above 3 kHz: a covered mic. EQ fixes it; Voice Isolation helps; EQ isn't scriptable. ## Captions - **Captions are for reading, not a court transcript.** "pro-product", "es--", "Hããã" were flagged as errors. Clean them. - **Correct written language.** The creator speaks slang; captions use the correct forms. Fix grammar, keep the voice. - **Your own filters can break words.** A filler rule removed the article "um" everywhere. Read every cue. - **Resolve's subtitle import shifts everything to the first cue.** Start the file at 00:00. ## Overlays - **16:9 layouts cover the face in 4:3.** Scale panels into the free strip; shrink full-frame type. - **Overlays and captions fight for the bottom.** Reserve the caption band in the kit. - **No lone "." on a line.** Glue punctuation to the highlight; pretty text wrapping. - **Get the story right.** The outro said "see you in a new country" while the creator was flying back home. Ask when unsure; don't name places they want to keep secret. - **Re-render under a new name.** Same-name files can play stale cached frames. ## Resolve API - `Timeline.ApplyGradeFromDRX` fails in 21; the per-item node graph works. - `AppendToTimeline` end is exclusive; audio-only clips count frames at their own rate; timecode counts 24 frames per second at 23.976. - Subtitles via the API land after the timeline end; a workaround crashed Resolve. - External scripts render only with `StartRendering([job], False)`, into a path without spaces. Restore the Deliver settings afterwards. ## Export - "Automatic" quality made 25 GB for 11 minutes. 50 Mb/s H.264 (or 30 Mb/s H.265) is plenty for YouTube. - Match colour tags to the project (Rec.709-A → "Same as Project") or YouTube looks faded on Apple devices.