cinematic-reel

Turn one raw landscape/cinematic clip (DJI Osmo Pocket, drone, phone) into a polished vertical 9:16 cinematic reel — scale/crop, color grade, beat-snapped chapter/title cards, optional ElevenLabs VO +

SKILL.md

GitHub

Cinematic Reel — scenic clip → vertical cut → all-socials schedule

End-to-end for a SINGLE scenic source clip (sunset, landscape, b-roll, drone, Osmo Pocket). Output: vertical 9:16 cinematic cut with title/chapter cards + optional VO/BGM, then optional multi-platform GHL scheduling. Single ffmpeg pass.

When to use

  • One scenic/cinematic source clip → want a vertical reel with grade + text cards.
  • Optionally: schedule it to FB/IG/TikTok/YT/LinkedIn/Pinterest at staggered times with per-platform copy.

Not for: talking-head/script reels (→ reel-studio), landscape screencast promos (→ demo-promo), static social cards (→ social-stack).

Pipeline

1. Probe + verify the arc (DO NOT skip)

ffprobe the source (res/fps/duration). Then extract frames across the timeline and eyeball them — card text must match what's actually on screen. Numeric/agent color analysis lies (auto-exposure, artificial lights). See references/pipeline-notes.md.

ffmpeg -ss <t> -i SRC -frames:v 1 -q:v 3 frame_<t>.jpg

2. Write the config + render (one pass)

Build a reel.config.json (schema in scripts/build_cut.py header). Snap card t_in to the BGM beat grid; card at VO handoff, one before the ~50% drop-off, ~50-60s cadence. Then:

python -u <skill>/scripts/build_cut.py reel.config.json

encoder:"auto" tries h264_nvenc and falls back to libx264 if the GPU driver is too old. Audio comes out 48kHz stereo, loudnorm ~-14.

3. QA the render

Probe streams (expect h264 WxH, aac 48000 stereo). Re-extract frames at each card t_in and confirm the overlay fired over the right image. Check loudness: ffmpeg -i OUT -af ebur128=framelog=quiet -f null -.

4. (optional) Schedule to all socials via GHL MCP

  1. mcp__ghl__locations_get-location → confirm timezone (e.g. Asia/Karachi +05:00).
  2. mcp__ghl__social-media-posting_get-account → account IDs (cached in references/ghl-accounts.md).
  3. Upload the rendered mp4 to GHL media → get the CDN URL (user does this, or media API).
  4. Write schedule.config.json (one entry per platform, each with its OWN human caption + hashtags), then python <skill>/scripts/make_schedule_plan.py schedule.config.json to get the exact payloads.
  5. Fire each payload with mcp__ghl__social-media-posting_create-post. Same-platform accounts batch in one call; different captions = separate calls.

Captions (per platform, human, psychology-led)

Different hook + different hashtag set per platform. Humanizer rules: no em-dash spam, contractions, varied sentence length, real specifics. Travel/personal = NO agency pitch; handle @yourhandle. Hooks by platform: FB = nostalgia + reciprocity question · IG = identity/FOMO + save-bait · TikTok = POV + curiosity gap · YouTube = curiosity + completion · LinkedIn = reflection + universal lesson.

Files

  • scripts/build_cut.py — single-pass render, config-driven, NVENC auto-fallback.
  • scripts/make_schedule_plan.py — offline builder for the staggered GHL create-post payloads.
  • references/pipeline-notes.md — the gotchas (read before debugging).
  • references/ghl-accounts.md — your account ID map + per-platform create-post rules.
  • assets/fonts/ — impact.ttf, arialbd.ttf.

Reference build

<repo>/sunset-reel-2026-05-27/ (60s + 4min cuts, PUBLISH-PACKAGE.md).