Video Studio: this skill lives in engines/legacy-editor. Before running anything from it, cd engines/legacy-editor and use that folder's copy (engines/legacy-editor/.claude/skills/clipper/), which has its scripts and reference files. Every path below is relative to engines/legacy-editor. This copy only lets Claude find the skill from the Video Studio folder.

Clipper: harvest short-form clips from a finished long-form

A second entry path, not a pipeline step. The main pipeline runs raw footage forward to one finished video. This skill runs the other direction: one FINISHED long-form video comes in, several short-form clips come out. Selection is the craft here; the build reuses the locked engine end to end (rough-cut splice, face-frame reframe, tiktok-raw captions), so a clip is a real job folder with the full audio chain, not a dumb trim.

This skill selects and builds clips. It does not edit the long-form. No re-cutting of the source, no graphics passes, no thumbnails.

Craft principle. A clip is not an excerpt, it is a standalone video that happens to be quarried from a longer one. If it needs the long-form's context to land, it is not a clip. Cold-open on the strongest sentence, never on wind-up.

---

Inputs

1. The source video: the PUBLISHED long-form (the file people actually saw), or any long talking video. Intake it like any job: create projects/<video-name>-clips/ and put the source in raw/ (hardlink when the file already lives in this repo, copy when external; never move the original). Always clip from the published file, not from a parent job's raw footage: an edited video's raw transcript does not match the published timeline.
2. Clip count and targets, if given. Default: 3 to 6 candidates per 10 minutes of source, 15 to 60 seconds each (90s hard cap).

Step 1: transcribe the source (once)

bash .claude/skills/rough-cut/scripts/transcribe.sh projects/<parent>

Skips automatically when transcript/words.json already exists (--force to redo). A long video takes a while on WhisperX large-v3: that is fine, accuracy over speed, and it only happens once.

Step 2: select the clips (the actual work)

Read the transcript and hunt for moments that stand alone. For each candidate, judge:

a. The hook. The first spoken line must grab cold, with zero context. If the strongest line sits mid-moment, open there and let the explanation follow it. b. Self-containment. Kill or cut around anything that leans on the long-form: "as I said earlier", "in this video", "back in step two". If the reference is load-bearing, drop the candidate. c. One idea, one payoff. A clip earns its length by paying off the hook. No payoff, no clip. d. Internal tightening. A clip's segments do not have to be contiguous: splice out filler, dead air, and detours inside the span exactly like a rough cut. Repeated line across takes: keep the LAST take.

Write two artifacts in the parent job folder (reviewed artifacts, not scratch):

• projects/<parent>/clips-plan.json: the machine plan make-clips.py builds from.
• projects/<parent>/clips-plan.md: the cut sheet for review, a table of # · name · source time · dur · hook line · why it stands alone, with ⚠️ on any candidate you are unsure about.

Report back tight: source duration, candidates found, and the 2 or 3 strongest hooks. Then stop: the plan gets reviewed before anything renders. Adjustments happen on the plan, then build.

clips-plan.json schema

{
  "job": "my-video-clips",
  "source": "raw/my-video.mp4",
  "clips": [
    {
      "id": 1,
      "name": "agents-run-while-you-sleep",
      "why": "strongest claim in the video, stands alone, natural 40s arc",
      "hook_text": "your editor works\nwhile you sleep",
      "hook_end": 2.8,
      "duration": 41.3,
      "segments": [
        { "start": 312.44, "end": 331.10, "transcript": "the hook and setup words kept here" },
        { "start": 338.72, "end": 361.40, "transcript": "the payoff, filler between spans cut" }
      ]
    }
  ]
}

Rules: segments use the exact cuts.json contract (SOURCE-file seconds, xfade and no_refine overrides allowed; clip is filled in automatically). hook_end is in CLIP-timeline seconds (sum segment durations up to the hook's last word). hook_text is the on-screen card in the creator's voice, punchy and lowercase-casual, 2 lines max; omit it (or leave empty) for no card, captions still run. name is a kebab content title, unique across the whole repo (clip folders share the /tmp/video-editor/<name>/ scratch namespace and the ~/Downloads export namespace with top-level jobs).

Step 3: build the approved clips

python3 .claude/skills/clipper/scripts/make-clips.py projects/<parent> --build

(On Windows run it as python, not python3 — Windows has no python3 name, only a fake Store stub.)

(--only <name> rebuilds one clip; no --build scaffolds the folders without rendering.)

Per clip this scaffolds a standard job folder projects/<parent>/clips/<name>/ (hardlinked source in raw/, parent words.json, authored cuts durable in transcript/cuts.authored.json plus the /tmp copy splice reads) and then runs the locked lane:

1. Splice: the real rough-cut engine, so refine, frame-snap, the static audio chain, and audio-qa.py all apply. Heed audio-qa's ⚠ lines: on limiter pressure re-run that clip with AMPLIFY_DB=8 ... --only <name>.
2. Reframe: face-frame.py measures the face and produces the face-centered 1080x1920 crop (center crop fallback when no face is detectable, e.g. screencast sections). The 16:9 splice is kept as outputs/<name>-169.mp4; the 9:16 crop becomes the clean base outputs/<name>.mp4 (audio stream-copied, never re-encoded). The explainer --verify standard does not apply here: that check is for the split-frame seam layout, not full-frame clips.
3. Captions: the LOCKED tiktok-raw preset (captions always on, hook card only when the plan gives hook_text) writes the deliverable outputs/<name>.final.mp4 directly, plus an export copy in ~/Downloads/ (VE_EXPORT_DIR overrides). finalize.sh is NOT used: it does not support nested job paths, and the builder already produces the canonical deliverable.

Then review like a second pass: watch the clips, adjust the plan (segments, hook text, drop a clip), and rebuild only what changed with --only.

Format note

Clips default to the TikTok/raw look, which is what a talking-head pull-quote wants. For a clip that deserves the full explainer treatment (top-half graphics, centered captions), treat its clip folder as a normal job and run pipeline steps 3 to 5 on it instead of step 3 above; it is a standard job folder, so every downstream skill works on it unchanged.

Gotchas

• Clip names must be repo-unique. /tmp/video-editor/ scratch and the Downloads export are keyed by basename alone; a clip named like an existing job collides.
• Plan times are SOURCE seconds; hook_end is CLIP seconds. Mixing the two puts the hook card over the wrong words.
• A rebuild regenerates the clip from the plan. Hand edits inside a clip folder do not survive --build for that clip; fold changes into clips-plan.json instead.
• Screencast-heavy clips get a center crop. If the interesting pixels sit off-center, add a "crop_x" note to the plan's why and adjust after the draft, or give the clip the explainer treatment instead.

Handoff

Each clip is a standard job folder. background-music, graphics-plan, or a Premiere/CapCut/Resolve replay can run on projects/<parent>/clips/<name>/ exactly as on any job. Posting is out of scope, as everywhere in this repo.

This skill does ONE thing: turn a finished long-form into reviewed, built short-form clips. It never edits the source video.