Herk Edit Kit by Nate Herk: edit videos, reels, and YouTube Shorts with Codex or Claude Code. 15 skills, transcript-driven cuts, storytelling, 406 motion-graphics cards, and HyperFrames templates.
From recording to finished edit
1. Create a project and transcribe
From the repository root run npm run new-video -- my-video. Put your recording
at video-projects/my-video/assets/raw.mp4. Preserve this original.
node scripts/transcribe-elevenlabs.mjs video-projects/my-video/assets/raw.mp4
This is Nate's default and uploads audio to ElevenLabs using your key and credits.
You may choose OpenAI Whisper, local Whisper, or another provider instead; see
provider setup and transcript normalization.
You can also supply an existing word-level
JSON transcript with words: [{text, start, end}] and audio_duration_secs.
2. Remove dead air
node .agents/skills/cut-silences/scripts/cut-silences.mjs video-projects/my-video/assets/raw.json --video video-projects/my-video/assets/raw.mp4 --out-dir video-projects/my-video/assets --output video-projects/my-video/assets/silenced.mp4 --apply
The retimed transcript is raw.silence-transcript.json. Tune gap and breathing
parameters in the skill; do not remove pauses that carry meaning.
3. Review mistakes
node .agents/skills/cut-mistakes/scripts/find-cut-candidates.mjs video-projects/my-video/assets/raw.silence-transcript.json --out-dir video-projects/my-video/assets
Read each proposal against the speech. Write reviewed decisions to
approved-cuts.json using { "cuts": [{ "start": 1.2, "end": 1.5, "reason": "stutter" }] }.
These times are on the SILENCED timeline. Empty cuts are valid.
node .agents/skills/cut-mistakes/scripts/apply-cuts.mjs video-projects/my-video/assets/raw.silence-transcript.json --cuts video-projects/my-video/assets/approved-cuts.json --video video-projects/my-video/assets/silenced.mp4 --out-dir video-projects/my-video/assets --output video-projects/my-video/assets/clean.mp4 --apply
Use raw.mistakes-transcript.json alongside clean.mp4. Copy that transcript to
the project's assets/transcript.json for beat validation. Never use raw timings
against the cleaned video. Claude users can substitute .claude for .agents.
4. Design the visual layer
Read video-storytelling and the motion philosophy. Fill a beat sheet with the spoken anchor, on-screen idea, visual, entry time, and callback. Pick a style and copy cards with their CSS/assets into your composition. Preserve the speaker's face and use a non-timed wrapper for camera reframing. See style-library skill. The graph-paper template is a background; it needs your story content on top.
5. Check, preview, and render
node scripts/preflight.mjs video-projects/my-video
node scripts/validate-beat-sync.mjs video-projects/my-video
cd video-projects/my-video
npx hyperframes lint
npx hyperframes preview
Run beat-sync only for anchored external sub-compositions; it fails when none are present. For inline compositions, check timing manually. Review Studio, then render a draft. Inspect encoded frames at hero moments and transitions, and listen at joins. Fix problems and write VERIFY.md before a standard render. Publication is a separate action requiring the student's instruction.