compose_video
pipeline as long explainers, with tighter scene counts and optional 9:16 output.
Defaults that work well for shorts: 4–5 scenes, 30–90s total, --no-text
(voiceover only), hybrid B-roll or --ai-images-only.
Quick recipe (LLM script)
--ai-images-only when you want Gemini/Pollinations images instead of stock photos
(requires GEMINI_API_KEY or POLLINATIONS_API_KEY).
Terminal tool short (custom script + CTA)
For repeatable shorts (intro → value → install tip → like/subscribe), use a JSON script instead of--llm. Keep narration punchy; last scene is your CTA.
terminal-short-script.json:
media_type per scene when you want explicit control (video for motion B-roll,
image for diagrams or UI stills). See Compose video.
Vertical 9:16 (Shorts / Reels)
compose_video has no --orientation flag yet. Set env vars for portrait output
and vertical stock search:
ai_video with VIDEO_ASPECT=9:16:
Useful flags & env
Natural language also routes correctly: