An agent skill that helps plan video edits through iterative dialogue — transcribe audio, extract key frames on demand, analyze visuals, and produce a structured edit plan.

Repository: https://git.nite07.com/nite/video-edit-planner-skill

Features

Audio transcription — bundled funasr-script with Fun-ASR-Nano (high quality, Chinese-optimized) and SenseVoice (fast preview) models
On-demand frame extraction — ffmpeg-based clip cutting (-c copy) + scene-change detection + uniform sampling, with hardware acceleration (CUDA/NVDEC preferred)
Artifact indexing — SQLite database tracks all transcriptions, clips, and frames to avoid duplicate processing across sessions
Vision analysis guidance — binary-search-style frame sampling strategy; works with any vision-capable model the agent's runtime provides
Iterative edit plans — Markdown-table output (timecodes, segment descriptions, actions, transitions, notes) refined through follow-up questions
Agent-agnostic — no platform-specific tool names; works with any agent framework (Hermes, Claude Code, Codex, etc.)

Quick Start

Prerequisites

Dependency	Install
`ffmpeg` + `ffprobe`	Linux: `pacman -S ffmpeg` / `apt install ffmpeg`; macOS: `brew install ffmpeg`; Windows: `winget install Gyan.FFmpeg`
`uv`	Linux: `pacman -S uv`; macOS: `brew install uv`; Windows: `winget install astral-sh.uv`; fallback: `pip install uv`
`python3`	Linux: `pacman -S python`; macOS: `brew install python`; Windows: `winget install Python.Python.3`

Install the skill

Copy or clone this repository into your agent's skills directory. For Hermes Agent:

git clone mygit:nite/video-edit-planner-skill.git ~/.hermes/skills/media/video-edit-planner

First use

The transcription scripts need a one-time dependency setup:

cd ~/.hermes/skills/media/video-edit-planner/scripts/transcription && uv sync

This creates an isolated venv with funasr, torch, torchaudio. If you've used these packages via uv before, the cache is reused — first-time cost is only the link step.

Workflow

1. Check dependencies (ffmpeg, uv, python3)
2. Gather inputs (video path + audio track index)
3. Ask editing requirements
4. Transcribe audio (skip if cached in index)
5. Extract clips & frames on demand (agent decides when transcript is insufficient)
6. Analyze frames with vision model (binary-search-style sampling)
7. Produce Markdown-table edit plan + overall recommendation
8. Iterate — user asks follow-ups, plan is refined

Project Structure

video-edit-planner/
├── SKILL.md                          # Skill definition (workflow, guidance, pitfalls)
├── README.md                         # This file (English)
├── README.zh.md                      # 中文说明
├── scripts/
│   ├── transcription/                # Bundled funasr-script (self-contained uv project)
│   │   ├── pyproject.toml
│   │   ├── uv.lock
│   │   ├── funasr_common.py          # Shared: ffprobe, audio extraction, model runner
│   │   ├── funasr_nano.py            # Fun-ASR-Nano entry point (high quality)
│   │   ├── funasr_fast.py            # SenseVoice entry point (fast preview)
│   │   └── funasr_regular.py         # Paraformer entry point (comparison)
│   ├── frames/
│   │   └── extract_frames.py         # Clip extraction + frame sampling (ffmpeg wrapper)
│   └── index/
│       └── manage_index.py           # SQLite index management (8 subcommands)
└── references/
    └── frame-extraction-guide.md     # Vision model token costs, resolution/batch guidance

Index Database

All processing artifacts are tracked in an SQLite database (<video_stem>.vedit.db) stored next to the video file:

transcription — JSON path, track index, duration
clips — start/end time, file path, extraction reason
frames — timestamp, file path, scene score, extraction method

This avoids re-transcribing or re-extracting frames for the same video across sessions.

License

MIT