Clip: long video → selected, polished shorts

Rules (modes, questions, honesty, data handling): ${CLAUDE_PLUGIN_ROOT}/references/rules.md. CLI below is ss = "${CLAUDE_PLUGIN_ROOT}/bin/shorts-studio" — always type the full path.

Input: $ARGUMENTS

0. Engine check

The commands below run the Shorts Studio engine. Installed as the Claude Code plugin it is already present and shorts-studio is on the Bash PATH. If shorts-studio doctor is not found (for example, this skill was added with npx skills add), ask the user before installing anything, then:

• Claude Code: /plugin marketplace add bloxy-studios/shorts-studio → /plugin install shorts-studio@shorts-studio.
• Other agents: git clone --depth 1 https://github.com/bloxy-studios/shorts-studio ~/.shorts-studio/engine and read every ${CLAUDE_PLUGIN_ROOT} path below as ~/.shorts-studio/engine.

Then run shorts-studio doctor --setup once (project-local runtime; never sudo or global installs).

1. Set up the project (ask only what is missing)

If the user has not specified them, ask ONE AskUserQuestion batch (≤ 3 questions): platform (YouTube Shorts / TikTok / Reels / generic), clip count + duration band (15–30 / 30–60 / 60–90 s), and editing preset (Social Dynamic / Clean Podcast / Comedy / Educational / Luxury Tech / Custom). Mode defaults to guided unless the user said otherwise. Autonomous: choose defaults (generic-9x16, 3 clips, 30–60 s, preset by content) and log them.

ss project init "<name>" --preset <id> --mode <guided|director|autonomous> --platform <p> --clips <n> --duration <min-max> [--language en|sw|auto]

2. Acquire and analyze (real audio + picture evidence)

ss ingest <project> "<url-or-path>"      # YouTube via yt-dlp; local files are hashed, probed, proxied
ss transcribe <project>                    # whisper.cpp, VAD-chunked; coverage verified vs speech activity
ss analyze <project>                       # scenes, audio events, face tracks, timestamped sheets

• A t= in the URL is a hint only; the full video is analyzed unless the user asked to start there.
• Acquisition failures are diagnosed (sign-in, private, geo, rate limit, outdated extractor). Never probe credentials; ask for an authorized local copy instead.
• Watch the sheets. analyze prints sheet paths (analysis/<source>/sheets/ov*.jpg). Read each image. For every reaction, gesture, on-screen text, screen demo, laughter or quality issue you actually see, record it: ss analyze note <project> --kind reaction --at 01:23.4 --detail "guest laughs at the joke" --confidence 0.8 --sheet ov01. This is what upgrades analysis from cv-only to agent-inspected; skip it and you must describe the result as unverified visually.

3. Find and rank moments

ss candidates <project> --count <n> --duration <min-max>
ss candidates preview <project> --top <n+2>

Read each candidate's preview sheet. Correct titles/hooks so they are truthful to the words spoken, reject weak or misleading ones, adjust boundaries if setup or payoff is missing, and attach what you saw: ss candidates note <project> --id c-003 --state rejected --reason "punchline depends on an earlier joke" ss candidates note <project> --id c-001 --title "..." --hook "..." --start 12:01.2 --end 12:44.0

If the source has fewer strong moments than requested, say so and offer the best ones — do not pad with weak material.

4. Let the user choose (guided / director)

Present a compact table: rank, time range, duration, title, why it holds, analysis level, confidence, preview path. Then:

1. Generate caption style previews on one representative excerpt: ss preview captions <project> --range <start-end> and show the comparison sheet (Read it, describe the six styles briefly).
2. Ask ONE AskUserQuestion batch: which clips (multiSelect), caption style (use the sheet you showed; recommend the preset default), editing intensity (restrained / balanced / energetic). Record answers: they become project decisions.

Autonomous: select the top non-overlapping candidates, preset default caption style, balanced.

5. Plan, render, inspect, repair

For each chosen candidate:

ss plan <project> --candidate <id> --style <caption-style> [--keyword "Product Name"]
ss render <project> --clip <clip-id>          # director mode: add --draft for the first one and ask to steer

render runs measured QA and writes report.md. Then inspect real frames: Read renders/<rid>/qa/contact.jpg and qa/cuts.jpg. Check framing, captions vs faces/UI, emphasis, cut smoothness, first/last frame. Record what you checked: ss qa note <project> --clip <clip-id> --render <rid> --what "contact + cuts sheets" --notes "...". Fix hard failures and visible problems with /shorts-studio:revise (bounded: ~3 repair passes, see project repair settings). If a hard failure remains, deliver it as a draft and say why.

6. Deliver

/shorts-studio:export packages final MP4s, covers, SRT/VTT/words.json, stems, plans and reports. Do not upload or post anything unless the user asked for that destination.

Known limits

• Visual analysis is sampled (faces at 2–5 fps, ~1 sheet frame per 10 s plus scene starts).
• Speaker↔face mapping relies on audiovisual evidence; when it is uncertain the framing falls back to two-person or wide layouts instead of guessing.
• Sources below 1080p are upscaled for 1080×1920 delivery; the report states the factor.

Completion check

For every delivered clip: render.json status done; qa.json verdict pass or pass-with-warnings (or explicitly reported as draft/blocked with reasons); an agent-visual inspection recorded; and you gave the user the MP4, report and preview paths.

Example

/shorts-studio:clip https://www.youtube.com/watch?v=VIDEO_ID Find five strong moments. Ask me to choose captions and editing intensity.