Captions in FilmCraft

House style: for any client edit, load edit-style (and its talking-head-edit or voiceover-edit sub-skill) first. It holds the approved principles and numbers this tool applies.

The user edits and exports in FilmCraft, so captions go into the project as text clips they can see, nudge and export themselves, not burned into a render. Tool: scripts/fc_captions.py (needs filmcraft-cli, ffmpeg, numpy, onnx-asr and faster-whisper; it imports the transcribe skill's Parakeet helper, so install that skill alongside, and the fonts come from the captions skill's fonts/).

The rules (a client's review, now the standard for every video)

1. Picture, caption and voice change on the same frame. Every cut on V1 starts a caption, and every caption starts on the frame its first word is heard: never a few frames late, never before the cut.
2. Blocks are whole script phrases, at most 4 words. Three words when that reads well, four only where the phrase needs it ("dimples that barely show," / "even in bathroom light.", not "dimples that barely" / "show, even"). Never split a name, a number and its unit, or "65% off".
3. Back to back while the voice runs, gone in silence. Each caption stays up until the next starts, unless the voice pauses for 0.4 s or more first: then it ends 2 frames after the voice stops. The last ends 2 frames after the last word. A caption is never shorter than half a second. --hide 0 puts them back to back everywhere.
4. A new picture appears only on a caption start or a V1 cut. Every B-roll clip, box still or full-screen insert starts and ends on the frame a caption starts or the main video cuts, never mid-caption, and no picture shows for under 8 frames. The Aleo video-3 client: "every time a new image comes to the screen its either when a new text starts or when the cut is happening of the main video. If you dont do this, it looks super messy" and called the misses "very rookie mistakes". Inserts placed on word times alone (start on the words they show) still landed 3–20 frames inside a caption on about a third of the changes.

Why the timing works this way