kevinbadi/social-agents/vertical-video-thumbnail

Vertical video thumbnail

9:16 (1080x1920) cover from the edited video. Run from the workspace root.

S=social-agents/skills/agent-posts/scripts/vertical-video-thumbnail/scripts
node $S/generate.mjs --video in.mp4 --out out/cover.png --caption-only          # transcript + draft only
node $S/generate.mjs --video in.mp4 --out out/cover.png \
  --transcript out/narration.json --keyword X --line1 "KICKER" --line2 "PAYOFF"  # the cover
node $S/generate.mjs --video in.mp4 --out out/cover.png --scene-grid out/cover-scene-grid.png \
  --line1 "NEW" --line2 "HOOK"                                                   # re-burn text only

Writes beside --out: cover.png, narration.json (transcript), caption.txt and caption-threads.txt (heuristic drafts, the agent rewrites them from BRAND.md), and a .json sidecar. --no-cta for videos without a spoken keyword.

Cover

Two modes, chosen automatically:

• Real frame (default). A frame from the video, fitted, with the two hook lines burned onto the bottom plate. Needs only ffmpeg and a bold font.
• Generated scene. Needs FAL_KEY, FAL_ALLOW=video-thumbnail, and two or more photos of the creator in social-agents/assets/identity/ (AGENT_POSTS_IDENTITY_DIR overrides). An optional identity.txt there describes the person for the prompt. assets/style-ref.png drives layout only. A locked or failing fal account falls back to the real frame; the cover step never fails the post.

Grid crop

Instagram Reels and TikTok profile grids are 3:4, center-cropped out of 9:16 (a 1080x1440 window at y=240..1680). That window is the photo: face, logos, background, no type. The hook lives only in the bottom 240px the grid crops away. In ffmpeg drawbox h is box height, so y=ih-240; drawtext uses y=H-186 / y=H-88.

Poses (generated scenes)

The pose bank in generate.mjs rotates by a hash of the video filename so retries match but every video differs; --pose "..." overrides. Never a chin-rest or thinker pose, never the pointing-at-camera grin, never prayer hands. No on-screen text, watermarks, view counts, or UI chrome in generated pixels. No em or en dashes in overlay copy.

Transcript

Local whisper first (WHISPER_MODEL, default base.en), fal whisper as fallback when FAL_KEY is set. Font: Arial Black, else DejaVu/Liberation Sans Bold.