SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsautoclip

autoclip

Design & creation

View repository

Updated 10-29 04:35

About this skill

AutoClip — Local-First AI Video Clipper

autoclip is an open-source, local-first AI video clipper that converts long-form video (podcasts, interviews, webinars, conference talks) into platform-ready 9:16 vertical clips with speaker-tracked reframing and animated captions.

No cloud uploads, no watermarks, and no subscription fees. Only transcript text is sent to an LLM for highlight detection; with a local Ollama model, zero data leaves the machine.

When to use this skill

  • Converting long YouTube videos, podcasts, webinars, or raw footage into vertical 9:16 clips for Shorts, Reels, or TikTok.
  • Finding viral hooks and high-engagement moments in long audio/video files automatically.
  • Reframing widescreen 16:9 footage into vertical 9:16 while tracking who is speaking using face detection (MediaPipe).
  • Burning animated karaoke or bold pop captions directly into the output video without external tools.
  • Operating the full pipeline locally with offline Whisper and Ollama for privacy or zero API costs.
  • Managing clipping jobs via CLI (autoclip clip, autoclip jobs) or reviewing clips in the local web app (autoclip serve).

When NOT to use this skill

  • Conversational raw footage editing (cutting stumbles, color grading, overlay animations) → use video-use.
  • Generating short videos from scratch using prompts or stock footage without long-form source footage → use moneyprinter-turbo.
  • Code-first programmable video compositions with Remotion → use video-production.
  • Offline video file compression, trimming, or format conversion without AI clipping → use compresso.
  • E-commerce / affiliate shopping shorts requiring platform compliance disclosure → use shopping-shorts.

Instructions

Step 1: Preflight environment with doctor

Before running any clipping jobs, verify host prerequisites:

autoclip doctor

Check that:

  • Python is 3.11 or 3.12 (Python 3.13 is unsupported by MediaPipe).
  • FFmpeg includes libass for subtitle burning (brew install ffmpeg-full on macOS, not plain ffmpeg).
  • Both ffmpeg and ffprobe are reachable on PATH.

Step 2: Configure LLM highlight detection

AutoClip requires an LLM to rank transcript highlights. Choose either local or API:

Option A: Local offline (Ollama)

ollama pull llama3.1:8b

Option B: API key (Stored in OS Keyring)

autoclip config set-secret anthropic
# or: autoclip config set-secret openai
# or: autoclip config set-secret gemini

Step 3: Run the clipping pipeline

Generate clips from a local video file or YouTube URL:

autoclip clip /path/to/long-podcast.mp4 --style bold_pop --aspect 9:16

Supported caption styles:

  • bold_pop: Chunky white text, dark outline, spoken word grows and turns yellow.
  • karaoke_fill: Words fill with accent color in sync with speech.
  • clean_lower: Minimal lower-third subtitle without bounce animation.
  • boxed: High-contrast text on solid background blocks.

All intermediate artifacts and exported clips are saved in ~/.autoclip/work/{job_id}/.

Step 4: Review and adjust clips

To inspect, preview, and trim clips visually:

autoclip serve

Open http://localhost:8000 to review candidate clips, adjust caption text or timing, and download final MP4s.

Examples

Example 1: Extract 3 Shorts from an interview video

User prompt:

Turn this 45-minute tech interview into 9:16 vertical clips with karaoke captions for YouTube Shorts.

Workflow:

  1. Run autoclip doctor to ensure ffmpeg-full and Whisper runtime are healthy.
  2. Ingest and clip: 
    autoclip clip ./interview.mp4 --style karaoke_fill --aspect 9:16
  3. Check generated jobs with autoclip jobs and verify exported MP4 files in ~/.autoclip/work/<job_id>/.
  4. Report clip titles, durations, and output paths to user.

Best practices

  • Ensure full FFmpeg build: Always check that FFmpeg has libass. If plain ffmpeg was installed on macOS via Homebrew, captions fail silently with No such filter: ass. Reinstall with brew install ffmpeg-full.
  • Match model to material: A small 7B model finds obvious topic breaks, while a frontier model (Claude 3.5 Sonnet, GPT-4o) spots subtle narrative hooks and jokes with significantly higher accuracy.
  • Shot-independent cropping: AutoClip detects scene cuts first and crops frames independently. Never force smooth camera pan interpolation across hard cuts.
  • Local storage cleanup: Long video jobs produce audio WAVs, frame crops, and temporary segments under ~/.autoclip/work/. Regularly clean up finished job folders after exporting.

References