muapi-youtube-shorts（muapi youtube 短裤）

YouTube Shorts Generator

End-to-end pipeline: Long Video → Transcript → Ranked Highlights → Vertical Clips.

Turns one long video into N viral-ready vertical mp4s. Each clip ships with a viral score (0–100), an opening hook line, and a one-sentence reason it should perform.

Reference implementation: https://github.com/SamurAIGPT/AI-Youtube-Shorts-Generator Underlying API: https://muapi.ai/playground/ai-clipping

Agent Execution Protocol

Step 1 — Collect Inputs

Ask once, then proceed:

If the user gave only a URL, use defaults and don't block on questions.

Step 2 — Verify Prerequisites

• muapi-cli installed and authed (muapi auth configure)
• ffmpeg on PATH (Whisper needs it for audio decoding)
• Python 3.10+ with openai-whisper installed (only if running the local transcribe stage)

If MUAPI_API_KEY is missing, stop and ask the user. Never invent a key.

Step 3 — Run the Pipeline

The standard path is the orchestrator script — it handles all eight stages in order:

bash library/social/youtube-shorts/scripts/run-youtube-shorts.sh \
  --source "<YOUTUBE_URL>" \
  --num-clips 5 \
  --aspect-ratio 9:16 \
  --whisper-model base \
  --output-json result.json \
  --view

The eight stages:

1. Download — pull the source video at the requested resolution (360/480/720/1080, default 720). For local files, skip.
2. Transcribe — local Whisper produces timestamped segments. Audio stays on the machine.
3. Classify content type — LLM tags the video (podcast / interview / tutorial / vlog / lecture / monologue) and density. Tunes the highlight prompt per type.
4. Chunk if long — videos > LONG_VIDEO_THRESHOLD (1800s default) are split into CHUNK_SIZE_SECONDS (1200s default) windows with CHUNK_OVERLAP_SECONDS (60s default) overlap so cross-boundary highlights aren't missed.
5. Rank highlights — LLM scans each chunk through VIRALITY_CRITERIA:

- Hook moments — strong opening line that stops the scroll - Emotional peaks — laughter, anger, vulnerability, awe - Opinion bombs — spicy, contrarian, debate-bait takes - Revelation moments — "wait, what?" reframes - Conflict — disagreement, tension, callouts - Quotable lines — tight, screenshot-worthy phrasing - Story peaks — climax of a narrative arc - Practical value — actionable insight a viewer will save Each candidate gets start_time, end_time, score 0–100, title, hook_sentence, virality_reason. Aim for 30–75s clips unless content dictates otherwise.

1. Dedupe — collapse overlaps. Rule: if two candidates overlap > 50%, keep the higher score, drop the other.
2. Top-N selection — sort surviving candidates by score, take num_clips.
3. Vertical auto-crop — render each highlight at aspect_ratio via muapi edit clipping. Auto-handles face tracking and screen recordings.

Quick Invocation Patterns

Single video, defaults:

bash scripts/run-youtube-shorts.sh --source "https://youtube.com/watch?v=VIDEO_ID"

Tuned for high-density podcast (more clips, larger Whisper model):

bash scripts/run-youtube-shorts.sh \
  --source "<URL>" --num-clips 8 --whisper-model medium --view

Square clips for Instagram feed:

bash scripts/run-youtube-shorts.sh \
  --source "<URL>" --aspect-ratio 1:1 --num-clips 3

Batch — urls.txt with one URL per line:

xargs -a urls.txt -I{} bash scripts/run-youtube-shorts.sh --source "{}"

Async submit (returns request_id, poll later):

REQUEST_ID=$(bash scripts/run-youtube-shorts.sh \
  --source "<URL>" --async --output-json - --jq '.request_id' | tr -d '"')
muapi predict wait "$REQUEST_ID" --download ./outputs

Platform Specs

Output Schema

{
  "source_video_url": "...",
  "transcript": { "duration": 1873.4, "segments": [...] },
  "highlights": [ /* every candidate, before top-N cut */ ],
  "shorts": [
    {
      "title": "The one mistake that cost me $50K",
      "start_time": 124.3,
      "end_time": 187.6,
      "score": 92,
      "hook_sentence": "Nobody talks about this, but it killed my first startup...",
      "virality_reason": "Opens with a number + regret, peaks on a contrarian lesson",
      "clip_url": "https://.../short_1.mp4"
    }
  ]
}

When reporting back to the user, surface for each clip: rank, score, time range, title, hook, and clip URL. Skip the raw transcript unless asked.

Tunable Knobs

Edit defaults inside the orchestrator or pass via flags:

Whisper Model Selection

• tiny / base — fast, English-leaning, fine for clean studio audio
• small / medium — better for accents and music beds
• large — highest accuracy, much slower; only worth it on a GPU

Pick base unless transcript quality is poor, then bump to medium.

Common Mistakes to Avoid

1. Skipping the dedupe step — without it, you ship near-duplicate clips that all came from the same hot moment.
2. Generic virality prompt — the highlight ranker must score against the eight signals above, not "interestingness."
3. Wrong aspect ratio for the platform — YouTube Shorts and TikTok are 9:16; LinkedIn often 16:9. Default to 9:16 only if the platform isn't specified.
4. Crop without face tracking — vertical crops on talking-head content must follow the speaker's face; static center-crop loses the subject.
5. Padding to hit num_clips — if dedupe leaves fewer survivors than requested, return what you have. Don't ship low-score filler.
6. Re-running the full pipeline on a 404'd clip URL — re-run only the crop stage for that highlight.

Failure Modes

• ffmpeg not found on PATH — stop and tell the user to install (brew install ffmpeg / apt install ffmpeg).
• Whisper produced no segments — likely no detectable speech or a hard language. Retry with --whisper-model medium --language <code> before declaring failure.
• API key missing or rejected — surface the exact error; don't fabricate a key.
• Job timed out — bump MUAPI_POLL_TIMEOUT and retry; don't silently truncate.
• Highlight ranker returned <num_clips — return what survived dedupe with a note.

Done Criteria

The skill is done when:

1. result.shorts has up to num_clips entries, each with a working clip_url.
2. The user has been shown the ranked list (score, time range, title, hook, URL).
3. If --output-json was set, the file exists and parses.