agentleFS
FSBuilder-mayuresh93FSBuilder-mayuresh93/product-film-director/skills/final-cut/SKILL.md

final-cut

Skill0 starsChanged 3 months agoMITView on GitHub

Cuts the finished film — conforms every shot to one master format, chains transitions, lays the voiceover, ducks the music bed, burns or ships captions, and exports the master plus vertical and silent cutdowns. Use when assembling a film from rendered shots, re-cutting after a script change, fixing audio levels or sync, or producing social and web versions of an existing film.

The file

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 3 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈3

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Add --preview while iterating — half resolution, fast encode, seconds instead of minutes.

    line 53
  2. Distinctive:

    Music must resolve, not stop: If the bed ends mid-phrase the film feels cancelled rather than finished.

    line 81
  3. Distinctive:

    If the user brings their own track instead, only use music they have the rights to.

    line 95
  4. Distinctive:

    Do not skip this. Check the actual file, not your intent:

    line 131

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Final cut

Everything is rendered. Now make it a film. One command does the assembly; the craft is in what you set before running it and what you fix after watching it.

Voice first, always

The edit is cut to the voice, not the other way round. If the VO is not laid yet, lay it now — every shot duration comes from it:

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/narrate.py" \
  --script script.json --outdir audio/ --align-shotlist shotlist.json

That writes audio/vo.wav, audio/captions.srt, audio/timing.json, and rewrites every shot's duration to speech + lead + tail.

It calibrates first. Piper reads at about 3.3 words/second — half again faster than anyone narrating a corporate film — so the whole script is measured and the pace corrected to --target-wps (2.5 by default) before any timing is trusted. Skip that and every shot is a third too short the moment a real voice replaces the scratch. The printed per-beat rate is then a writing signal: a beat still above 2.8 w/s is denser than the rest, and the fix is cutting words, never speeding the read.

Each shot's beat field says which line it serves. When several shots share one beat, that beat's screen time is divided between them in the proportions the shot designer set. Shots matching no beat keep their hand-set duration and are named in a warning — read it, because a shot that silently keeps a guessed duration is a shot that drifts.

Swapping in a real recording later: put it at audio/vo.wav and re-cut. If the read is close to the scratch timing the film barely moves. If it is not, re-time by hand: timing.json has every beat's start.

Voices: en_US-ryan-high (default, warm male), en_US-lessac-medium (neutral male), en_GB-alan-medium (British male), en_US-amy-medium (female). Pass with --voice. Listen to one beat before committing to the whole film.

Cut it

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/assemble.py" \
  --shotlist shotlist.json --shots-dir shots/ --out out/film.mp4

Add --preview while iterating — half resolution, fast encode, seconds instead of minutes. Drop it for the master.

The assembler conforms everything to one format (fps, size, SAR, pixel format), which is what stops generated 24fps footage and 30fps graphics from stuttering against each other. It also honours look.grade and per-shot grade, which is where generated and rendered material get pulled into one skin.

Encoding a full-length master takes a few minutes. In an agent sandbox, launch it detached rather than letting the shell time out:

nohup python3 "${CLAUDE_PLUGIN_ROOT}/scripts/assemble.py" \
  --shotlist shotlist.json --shots-dir shots/ --out out/film.mp4 > /tmp/cut.log 2>&1 &

Sound

Sound is half the film and gets a tenth of the attention. The defaults are already right; the judgement calls are these:

  • Music sits at about −22 dB under the voice and is side-chain ducked, so it breathes out of the way of each line and swells in the gaps. If you can follow the tune while someone is speaking, it is too loud.
  • The whole mix normalises to −16 LUFS, which is right for web and for most internal players. −14 if it is going to YouTube exclusively.
  • Give the film one moment with no voice. Two-thirds through, one beat of picture and music only. It is where the film stops being a presentation.
  • Music must resolve, not stop. If the bed ends mid-phrase the film feels cancelled rather than finished. Trim the music to land on the last fade.

Set it in shotlist.json:

"audio": {"vo": "audio/vo.wav", "music": "audio/bed.wav", "music_gain_db": -22}

No bed yet? Load score-film and generate one with Google Lyria. It writes the bed at the film's exact runtime with a resolved ending, so the assembler's loop never engages and the last phrase lands on the final fade — which is the "must resolve, not stop" rule handled for you rather than left to trimming.

If the user brings their own track instead, only use music they have the rights to. Ask; do not assume, and do not go and source it for them.

Captions

Ship them. Most internal video is watched muted the first time.

  • Separate file (default): audio/captions.srt alongside the mp4. Better for players that support it, and editable.
  • Burned in: --burn-captions. Necessary for social and for embedding where you do not control the player. Do it on the cutdown, not the master.

Deliverables

# master — 16:9, captions as a sidecar file
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/assemble.py" \
  --shotlist shotlist.json --shots-dir shots/ --out out/film.mp4

# vertical cutdown — 9:16, burned captions, for phones and social
python3 "${CLAUDE_PLUGIN_ROOT}/scripts/assemble.py" \
  --shotlist shotlist.json --shots-dir shots/ --out out/film-9x16.mp4 \
  --vertical --burn-captions

The vertical cut pads rather than crops, so nothing important is lost, but check it: a shot composed for 16:9 often has its subject at the edge. Shots that break in vertical should be re-framed in the shot list (pad_color, or a per-shot crop in grade) rather than left broken.

Also worth producing, if asked: a silent master for live presentation with a human speaker, and a 30-second cutdown built by disabling beats ("disabled": true) rather than deleting them — disabling keeps the reason the cut exists.

Verify before you send

Do not skip this. Check the actual file, not your intent:

ffprobe -v error -show_entries \
  format=duration,size:stream=codec_name,width,height,r_frame_rate,channels \
  -of default=nw=1 out/film.mp4

ffmpeg -i out/film.mp4 -af volumedetect -f null - 2>&1 | grep -E 'mean|max'
  • Duration within a second or two of the target
  • 1920×1080, yuv420p, h264, AAC 48kHz stereo
  • max_volume at or below −1.0 dB — anything at 0.0 is clipping
  • mean_volume around −18 to −20 dB
  • Then sheet the film and look at it: contact_sheet.py --film out/film.mp4 --out qc/

Finally: watch it once, all the way through, with sound. There are always three things wrong. Usually a beat that is too long, a transition that should be a cut, and one line that needs a word removed. Fix those three and it is done.

When something is off

  • Voice drifts out of sync — a shot duration was edited by hand after --align-shotlist. Re-run narrate with --align-shotlist and re-cut.
  • A join stutters — one shot's fps or SAR is odd; the conform handles it, but check the source clip actually decoded.
  • Film is too long — cut a whole beat. Never speed up the voice; the difference between 2.5 and 2.9 w/s is the difference between authority and anxiety.
  • Film feels flat — usually every join is a fade. Change most of them to cuts.
  • ffmpeg failed — the full log is at /tmp/assemble-ffmpeg.log.

References

  • references/audio.md — VO direction, music selection, ducking, loudness targets and how to time a real recording against the scratch track
  • references/deliverables.md — formats, platform specs, cutdowns, and what to hand over alongside the film

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.