agentleFS
PicsArtPicsArt/media-studio/skills/speech-aware-cutting/SKILL.md

speech-aware-cutting

Skill1 starsChanged 10 days agoMITView on GitHub

Choose the moments of talking footage worth keeping and cut them so no one is clipped mid-sentence - the transcript sets every in and out, the watch picks what is worth it - then read the cut as the viewer will hear it with scripts/review_cuts.py, which flags chopped words, filler and seams that open on "but" or "she", with the time to move each edge to. Use when cutting a short, a reel or highlights from an interview, podcast, vlog, talk or any footage where people speak, and when a finished cut sounds choppy.

1 other project uses a version of this, 1 of them edited.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

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

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈0

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

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

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

    Read every source as the footage-analysis skill says for talking footage: a transcript kept as the tool returned it (transcript-<source …

    line 16
  2. Very specific:

    Write the plan as the reel's clip list, the one the footage-reframing skill's scripts/reel_cut.py reads, with each spoken source's …

    line 40
  3. Very specific:

    Run python3 scripts/review_cuts.py reel.json on that same file.

    line 45
  4. Very specific:

    Read the script yourself too, as prose, for what is clunky.

    line 63

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 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

Speech cuts

A short cut from talking footage lives on its words. The framing can be perfect and the short is still unwatchable when it opens on half a sentence or spends its first two seconds on "yeah, no, totally". So the cut is planned from what is said, and checked as what is heard, before anything is built.

Read the footage

Read every source as the footage-analysis skill says for talking footage: a transcript kept as the tool returned it (transcript-<source id>.json where you run commands), which sets every boundary, and a watch, which picks the moments. A photo has no speech at all: it is a beat, not a backbone.

Plan the cut

  1. One pool of moments. Take the watch's candidate moments from every source into one list, each with the source it came from: its times count on that source's clock, as the footage-reframing skill says.
  2. Snap each to whole thoughts. A clip's in is a sentence's first word, or a word after a real pause; its out is a sentence's last word, or a word before one. The transcript's cutPoints mark the clean ends.
  3. Order and trim to the target. Order the moments by their loops, as the retention-hooks skill says. Aim for the target length within about 15%. When the length and a clean sentence conflict, the sentence wins and the length moves. Never cut a word out of a sentence to save time: the ear hears the splice at once, where the eye would forgive a picture cut.
  4. Place photos at sentence edges, never inside one, each held long enough to read.
  5. Keep what the user saw work. A recut or a fix starts from the last version the user saw, never from zero. Every moment they liked stays, with the picture it had; a fix for a constraint (a clipped word, a pause, the length) moves an edge, never swaps the moment for another. A moment dropped, moved or given a new picture gets one line in the message to the user. Where the use case keeps the edit-pipeline record, its keepers ledger and scripts/meaning_diff.py do this, as the edit-pipeline skill says.

Write the plan as the reel's clip list, the one the footage-reframing skill's scripts/reel_cut.py reads, with each spoken source's transcript beside its other fields (its file path, where you run commands).

Review it as heard

Run python3 scripts/review_cuts.py reel.json on that same file. It joins each clip's words in play order into the heard script, and checks three things per clip:

CheckFails whenIts repair
Integrityan edge cuts through a word, or the clip opens or stops mid-sentencemove that edge to the time given
Substancemost of the clip is filler; the opening clip is held to a stricter sharedrop it, or take the sentence beside it
Fluencya clip opens on a connective ("but", "so") with nothing before it, or the reel opens on a pronoun no one has been named formove the in point back to what it refers to, or lead with another clip
Contexta clip opens on an answer ("Yes", "Exactly") whose question the cut dropsmove the in point back to the question

Each clip also lists the speakers its words carry. Use them when the footage-reframing skill asks who a clip should hold on.

The review also returns context: each name the reel says for the first time, each "she" or "that" right after a seam, each "as I said", and each seam that jumps far in its source, with the second of the reel it is heard at and, when the cut drops it, what the source said first. These are not faults. They are the places a listener who knows only the reel may get lost.

Read the script yourself too, as prose, for what is clunky; whether a clean clip is worth its seconds is your call.

Apply the repairs to the plan and run it again until it is clean. A repair changes numbers in a plan, so it costs nothing. The review never blocks the cut: if every candidate is thin, build from the strongest and say so in one line.

A clip it marks silent (a photo, a shot with no transcript) is a beat of silence in the heard script. A segment the transcript returned without word times is checked only at its sentence edges, and the review says so.

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.