Video Editor Agent

A Claude Code skill pack that edits short-form videos end to end: raw footage in, a finished edit out — style cloned from a reference reel, sound-designed with ElevenLabs, QA'd frame by frame and in dB, delivered on a live review page with timeline comments, and revised round after round until sign-off.

This is not a rendering library. It is a set of skills, local project tools and a QA engine that make a Claude Code session behave like a working video editor with a proven pipeline and a "verify pixels and dB, not intentions" culture. It gets better with every real edit: the process goes into the skills, the personal details stay in a gitignored context file.

The pipeline

footage + (optional) reference reel + brand/site + ratio
        │
        ▼
[0] video-edit-pipeline picks the lane
      talking head + brand ─────────────→ [1]–[7] below
      creator's own reel look ──────────→ reel-recut
      multi-take screen+camera+mic ─────→ arcads-video-edit
      event recap: takes + B-roll ──────→ recap-video
      one long file of many hooks ──────→ hook-splitter (→ hook-variations)
      AI-actor footage that sounds fake → ai-audio-sound-design
        │
        ▼
[1] reel-style-clone ──────────→ STYLE-GUIDE.md (only if a reference exists)
        │
        ▼
[2] branded-ad-edit
      ingest → two crop candidates (zoom vs blur-pad — show both)
      whisper transcript → storyboard from the style guide
      hand-authored composition + karaoke captions + SFX markers
      B-roll: broll-capture (real screens) / openart-broll / arcads-broll (generated)
        │
        ▼
[3] sound-design ──────────────→ ElevenLabs SFX kit + music bed, mixed by math
        │
        ▼
[4] QA loop (video-qa + tools/video-qa)
      hyperframes check → ONE multi-timestamp snapshot → LOOK with vision
      → fix → render MP4 → the engine verifies the RENDERED file (frames + dB + seams)
        │
        ▼
[5] video-review-canvas ───────→ live review URL (reply leads with the link)
        │
        ▼
[6] revision rounds
      read notes → screenshot the exact frame per note → fix
      edl-tighten for pacing → new file per version → republish same slug
      per-note QA table with evidence from the rendered file
      one cut → many variants: hook-variations, then naming-convention
        │
        ▼
[7] (optional) capcut-export ──→ CapCut draft for a human's final pass

Quickstart

1. Clone this repo and open it in Claude Code.
2. Run bash scripts/setup.sh --demo. It installs local QA dependencies, creates missing config templates and builds a synthetic edit, catalog, QA evidence and versioned review page. Requires Python 3.10+, Node.js 20+, ffmpeg/ffprobe; no paid APIs or model downloads. SETUP.md covers prerequisites, workflow-specific setup and safe skill linking.
3. Fill in your brand, defaults and projects directory in MASTER_CONTEXT.md (where the videos live — default outputs/). Configure keys only for workflows you use.
4. Drop your raw footage into the projects directory (or footage/) and say: "edit this like <reference reel>" — or just "edit this video". The video-edit-pipeline skill takes it from there.

Project and review tools

• Review canvas v2: version picker, synchronized comparison, open/resolved notes, replies and before/after frame evidence. Shared reviews use append-only here.now records; an explicitly local mode supports demos without hosting. Export the complete revision ledger as JSON and import it into project state.
• Storyboard coverage: qa:storyboard checks planned IDs and timed HTML intervals, then extracts contact sheets from the actual MP4. Static coverage and visual verification are reported separately; see the QA engine.
• Resumable projects: python3 tools/editor/editor.py project init|status|resume records source identities, stage artifacts, renders, approvals and outstanding notes. Changed or missing files invalidate the relevant progress. Explicit configured commands can resume one stage at a time.
• Reusable footage catalog: python3 tools/editor/editor.py catalog ingest|search indexes media metadata, thumbnails, local transcripts, selected intervals and past uses. Stable content IDs prevent duplicate indexing. See project tools for full commands and transcript options.
• Offline regression suite: bash scripts/test.sh --integration covers production failures, project/catalog behavior, setup safety, the review model and a real media demo. GitHub Actions runs the same suite on Linux. Paid/live transcription is opt-in.

QA reports include layer coverage; unavailable checks are not labeled a clean pass. Review instructions and sampling configuration participate in cache identity. The recut renderer uses bounded lossless batches and one final AAC encode for long cut timelines. Serve any local review with python3 scripts/serve-review.py --directory <review-folder>; it supports the byte-range requests video players need for frame seeking.

Using it from your own working repo

Most people keep their videos, queues and brand tooling in a repo of their own. Keep that, and point it at this pack instead of copying skills into it:

• Link the skills with bash "<pack>/scripts/setup.sh" --no-install --link-skills "<working-repo>". Setup computes relative links for the actual directory layout and preserves existing paths. A session started in your repo loads the skills; their files stay here.
• Set the projects directory in MASTER_CONTEXT.md (and VIDEO_PROJECTS_DIR in .env) to your media folder. Media never moves into this repo.
• Alias the QA engine in your package.json: "qa:video": "npm --prefix \"<path to this repo>/tools/video-qa\" run qa:video --". Relative paths resolve from where you run it; your .env is read first.
• Improvements from every real edit go back into the skills here — generic process only. Your client names, fees, reviewer preferences and machine paths belong in MASTER_CONTEXT.md, which never leaves your disk.

Prerequisites

External dependencies are documented, not vendored — nothing in this repo ships a copy of HyperFrames, whisper models, or ffmpeg.

Skills catalog

Limitations (honest)

• CapCut export is work-in-progress. The layered export works but the schema patch is still being hardened — always verify the draft opens in CapCut before relying on it.
• Shared canvas delivery requires here.now. Local mode runs without hosting and keeps feedback in that browser; export its JSON to transfer notes. Shared mode persists feedback through the existing public review link. Local mode is not a multi-device review store.
• Transcription quality tracks your whisper model choice; tiny models miss words that then miss captions.
• The main build assumes single-subject talking-head source footage; multi-shot sources go through arcads-video-edit (screen + camera takes), recap-video (spoken takes and event B-roll), or hook-splitter (one long composite) first.
• Default tests are offline. The optional live transcription integration uses macOS say and an installed transcriber; deterministic seam tests run without it.