Conversation-driven video editing for AI agents. Transcribe raw footage with ElevenLabs Scribe, pack phrase-level transcripts, reason over an EDL, render with ffmpeg, and self-evaluate cut boundaries — all through a typed Node.js CLI.

video-use gives an LLM structured text (word-level transcripts) plus on-demand visual composites instead of dumping video frames. An agent reads takes_packed.md, proposes a cut strategy, writes edl.json, and renders final.mp4 into <videos_dir>/edit/.

flowchart LR
  subgraph inputs [Inputs]
    RAW[Raw takes]
    ENV[.env + ffmpeg]
  end

  subgraph pipeline [Pipeline]
    T[transcribe]
    P[pack]
    E[EDL reasoning]
    R[render]
    V[self-eval timeline]
  end

  subgraph outputs [Outputs]
    MD[takes_packed.md]
    MP4[final.mp4]
  end

  RAW --> T --> P --> MD
  MD --> E --> R --> MP4
  R --> V --> MP4
  ENV --> T
  ENV --> R

• Word-level Scribe transcription with speaker diarization, audio events, and per-file caching
• Phrase packing — silence-aware markdown optimized for LLM token efficiency
• EDL render pipeline — per-segment grade, 30 ms audio fades, HDR tone-map, overlay compositing, subtitle burn-in last
• Timeline composites — filmstrip + waveform PNGs for cut-point drill-down
• Optional Redis cache — cross-session transcript persistence when enabled
• Strict TypeScript — shared libraries, Vitest tests, ESLint flat config

git clone https://github.com/browser-use/video-use ~/Developer/video-use
cd ~/Developer/video-use
npm install
npm run build
cp .env.example .env
# Edit .env — set ELEVENLABS_API_KEY

Register the skill with your agent (symlink the repo into your skills directory). See install.md for agent-specific steps.

Set up https://github.com/browser-use/video-use for me.
Read install.md first, then SKILL.md for daily usage.
Use npx video-use for all editing commands.

Copy .env.example to .env:

Verify Redis connectivity:

npx video-use redis ping

npx video-use transcribe <video> [--edit-dir <dir>] [--num-speakers N]
npx video-use transcribe-batch <videos_dir> [--workers 4]
npx video-use pack --edit-dir <dir> [--silence-threshold 0.5]
npx video-use timeline <video> <start> <end> [-o out.png]
npx video-use render <edl.json> -o final.mp4 [--preview] [--build-subtitles]
npx video-use grade <input> -o output [--preset warm_cinematic]
npx video-use redis ping

Run npx video-use --help for full option lists.

video-use/
├── src/
│   ├── cli/           # Commander entry point
│   ├── config/        # env + logger
│   ├── lib/           # ffmpeg, scribe, pack, grade, render, timeline
│   ├── redis/         # connection manager + cache
│   └── types/         # shared TypeScript interfaces
├── tests/             # Vitest unit tests
├── docs/AUDIT.md      # internal architecture audit
├── SKILL.md           # agent editing instructions
├── install.md         # first-time setup guide
└── skills/manim-video/  # vendored animation skill

flowchart TB
  CLI[src/cli/index.ts]
  CFG[src/config]
  LIB[src/lib]
  REDIS[src/redis]
  TYPES[src/types]

  CLI --> CFG
  CLI --> LIB
  LIB --> CFG
  LIB --> TYPES
  LIB --> REDIS
  REDIS --> CFG

npm install
npm run dev -- --help          # run CLI via tsx without building
npm run typecheck              # tsc --noEmit (strict)
npm run lint                   # ESLint
npm run test                   # Vitest
npm run build                  # emit dist/
npm run validate               # all of the above

sequenceDiagram
  participant EDL as edl.json
  participant X as extract segments
  participant C as concat base
  participant O as overlays
  participant S as subtitles
  participant L as loudnorm

  EDL->>X: per-range grade + fades
  X->>C: lossless -c copy
  C->>O: PTS-shifted overlays
  O->>S: burn captions LAST
  S->>L: -14 LUFS normalize
  L->>EDL: final.mp4

npm run test

Tests cover phrase packing logic and Redis manager lifecycle (disabled-mode paths). Integration tests against ffmpeg and Scribe are intentionally manual — they require external services and burn API credits.

Ensure .env exists at the repo root with a non-empty key, or export the variable in your shell.

Confirm ffmpeg and ffprobe are on PATH and support libx264. Run ffprobe -version.

Set REDIS_ENABLED=true and confirm Redis is running locally (redis-cli ping → PONG).

Subtitles must be applied last in the filter chain (Hard Rule 1 in SKILL.md). Use npx video-use render — it enforces this order.

The render pipeline detects PQ/HLG sources and applies tone-mapping automatically. If issues persist, probe with ffprobe -show_entries stream=color_transfer.

Does the LLM watch the video? No. It reads packed transcripts (~12 KB per hour of takes) and requests timeline PNGs only at decision points.

Where do session outputs go? Always <videos_dir>/edit/ — never inside the video-use repo.

Can I skip Redis? Yes. Redis is off by default. Filesystem caching in edit/transcripts/ is always used.

What animation engines are supported? HyperFrames, Remotion, Manim, and PIL sequences — see SKILL.md. Each renders inside edit/animations/slot_<id>/.

How do I update the skill? git pull in the clone; re-run npm install && npm run build if dependencies changed.

1. Fork and clone the repository
2. Create a feature branch
3. Run npm run validate before opening a PR
4. Follow existing TypeScript conventions (strict mode, ESM imports with .js extensions)

See SKILL.md for production editing rules and docs/AUDIT.md for architecture notes.

MIT — see LICENSE.