A faceless Short is a vertical video with a voiceover, a headline and captions but no person on camera. faceless-shorts is an open-source (MIT) Agent Skill that turns a short scene script into a finished 1080x1920 MP4: each scene is voiced through the Miragea speech API, its length is set from the real audio, and word-highlight captions are burned in. It needs only Python and ffmpeg: no GPU, no stock-footage key, no second LLM key.
1. Write the script
{"scenes": [
{"id": "hook", "keyword": "Honey never spoils", "text": "Here is a food that never goes bad. Honey."},
{"id": "why", "keyword": "Why?", "text": "It is so dry and so acidic that bacteria simply cannot grow in it."},
{"id": "cta", "keyword": "Follow for more", "text": "Follow for one tiny fact every day."}
]}
2. Estimate, then make
git clone https://codeberg.org/miragea-voice/miragea-skills && cd miragea-skills
export MIRAGEA_API_KEY=... # create one at https://miragea.ai/workspace/api
python3 skills/faceless-shorts/scripts/shorts.py script.json --voice "Ava" --dry-run
python3 skills/faceless-shorts/scripts/shorts.py script.json --voice "Ava" --theme sunset -o short.mp4
Unchanged scenes are cached by hash, so editing one line only re-voices that line. No saved voice? Clone one from a public sample with --reference-audio-url and --reference-text.
Real result
A 17-second honey-facts Short made exactly this way. Speech recognition on the final MP4 returns the script word for word.
Use it from an agent
Install the skill collection in Claude Code, Codex or Cursor and ask: "Make a 30-second faceless Short about why the sky is blue." The agent writes the scenes and runs the script. Agent install line: Install https://miragea.ai/skills/miragea-audio/install.md.
FAQ
Does it need a GPU? No. Voice comes from an API; video is drawn by ffmpeg.
Can I use my own images? Yes, add "image": "photo.jpg" to a scene for a slow zoom.
Where is the code? https://codeberg.org/miragea-voice/miragea-skills (MIT, 82 offline tests).


Top comments (2)
Setting each scene's length from the real audio is the detail that makes this work; most pipelines I have seen guess durations and the captions drift by the third scene. Caching unchanged scenes by hash is a nice touch too. One thing I would add for anyone running this on a real channel: vary the visual layer per video, not just the script. If every Short shares the same theme, font and three-scene skeleton, it reads as a template to viewers and to YouTube's reused-content review. Even rotating the hook scene's background and caption style across a week helps. And if the voice is synthetic and realistic, ticking the altered-content box in Studio costs nothing.
장면별 음성을 해시로 캐시하고 실제 오디오 길이로 영상을 맞추는 흐름이 수정 비용을 줄이는 데 좋겠네요. 캐시 키에는 텍스트뿐 아니라 음성·속도·모델 설정도 포함되는지 궁금합니다. 같은 대사라도 발화 설정이 바뀌면 음성과 자막 타이밍을 함께 다시 계산해야 할 것 같습니다.