Summit Video Pipeline

summit-video-pipeline

End-to-end video editing pipeline driven by Claude Code + CLI tools only. No GUI editor in the loop.

When to use

• User has raw footage (one or more clips) and wants a finished short-form video.
• User wants filler removed (umms, ahs, dead air, retakes).
• User wants motion graphics, lower-thirds, captions, b-roll cuts, or animated overlays.
• User wants the entire edit driven by a single prompt or short brief.

If the task is purely a render of an existing edit, skip stages 1–4 and jump to stage 7.

Upstream stack (install once)

Prereqs: Node 20+, FFmpeg with libass (brew default lacks it — use brew tap homebrew-ffmpeg/ffmpeg && brew install homebrew-ffmpeg/ffmpeg/ffmpeg), whisper-cpp, headless Chrome/Chromium, ~5 GB free disk, 16 GB RAM recommended. Run npx hyperframes doctor after install. Verify libass: ffmpeg -filters | grep subtitles.

CLI tools used

• ffmpeg — cut, concat, mux, encode, scale, audio extract.
• whisper-cli (whisper-cpp) — transcript + word-level timestamps for cut points and subtitles.
• npx hyperframes — HTML→video renderer.
• manim (optional, via video-use skill) — programmatic 2D math/diagram animations.
• yt-dlp — pull reference footage if needed.
• imagemagick — thumbnail/poster generation.

Pipeline stages

Each stage lives in references/. Load only the stages needed for the current task.

System flow (full pipeline)

raw/                        # user drops footage here
  clip-01.mov
  clip-02.mov
  music.mp3 (optional)
brief.md                    # one paragraph: hook, audience, key points, length, vibe

1. Ingest — normalize fps/resolution/codec, write manifest.json.
2. Transcribe — whisper-cli over each clip → transcript.json w/ word timestamps.
3. Cut script — Claude reads transcript + brief, produces edl.json (in/out points per clip, ordering, intentional pauses).
4. Assemble — ffmpeg concat from EDL → rough-cut.mp4.
5. Motion graphics — Claude writes HTML/CSS/GSAP compositions in compositions/, renders each via npx hyperframes render → transparent-background MP4s.
6. Subtitles — whisper word timestamps → SRT, optional burn-in via ffmpeg subtitles= filter.
7. Composite + render — ffmpeg overlay compositions on rough cut, mux audio (music ducked under VO), encode H.264 9:16 1080×1920 → final.mp4.

Operator prompt template

Role: Edit this video as a Claude-Code-driven editor using summit-video-pipeline.
Context: [project, audience, platform — e.g. Summit IG Reels]
Objective: [one sentence, e.g. "60s reel teaching Freedom at 45 mini-app"]
Inputs: raw/ + brief.md
Constraints: [length, aspect, vibe, brand colors, no music / music track]
Output: final.mp4 + subtitle.srt + thumbnail.jpg

Single-stage shortcuts

• "transcribe this" → stage 2 only.
• "cut the umms out" → stages 2 + 3 + 4.
• "add a lower-third saying X at 0:12" → stage 5 only.
• "burn captions in" → stage 6 only.
• "render the final" → stage 7 only.

Quality gates

Before marking done:

• final.mp4 plays end-to-end with no black frames or audio dropouts.
• Subtitle drift ≤ 200ms vs spoken word at any point.
• Output matches target aspect (9:16 default) — verify ffprobe.
• No watermarks from upstream tools.
• File size reasonable for platform (< 100 MB for IG Reel).

Provenance

Reverse-engineered 2026-04-27 from @cooper.simson IG reel DXiyhqIjn_n. Cooper's caption: "What took 2 hours in Premiere and a team of experts now takes 10 mins." Cooper explicitly contrasts this with Remotion ("too limited"). HyperFrames is HeyGen's open-source successor approach — HTML/CSS/GSAP + headless Chromium.