# rea-video: break down how any video was made (JuneYaooo/rea-video) — skill card for AI agents

> AgentGid skill card. Checked on 2026-11-14 against commit 4089199 (version 0.1.0).
> Every fact comes from the pack's own files at that commit. Numbers marked "as reported" are the author's.
> Human page: https://agentgid.com/skills/juneyaooo/rea-video/

> Note: the pack has changed since this check; re-read its files before relying on details.

## Explain it to a non-programmer
Give your AI assistant a video and it works out, shot by shot, how it was made: filmed, AI-generated, 3D, animated, screen recording or built in code. It also lists the captions, stickers, voice, music and sound effects, and writes a plan to remake the video with your own material.

## Who it is for
- You saw a short video or ad and want to know if it is real footage or AI, and how it was made
- You want to remake a video in the same style with your own product, people or footage
- You use Claude Code, Codex or another assistant that can run commands on your computer
- Not for: You want a finished video in one click: it writes a breakdown and a plan, it does not render a video
- Not for: You want to repost someone else's video: the author says it is for learning and remaking with your own material

## What it is
- Kind: scripts, command
- License: Apache-2.0
- Repository: https://github.com/JuneYaooo/rea-video

## What it does
- Splits the video into shots and the sound into voice, music and effects
- Judges each shot as one of 11 types: live action, film clip, AI video, avatar, 3D, 2D animation, stop-motion, code template, screen recording, still image with pan or zoom, still frame
- For live footage: framing, camera position, lighting setup, colour grade and editing pace
- Lists the layers on top: timed captions, styled text, titles, stickers, picture-in-picture
- Marks real material you must supply yourself (screenshots, photos, recordings)
- Writes a remake plan: AI generation units with prompts, filming sessions with a gear list, or a Remotion code project
- An assistant that can see images reviews the low-confidence shots and merges its answers into the report
- Never cuts: Evidence for every conclusion; uncertain points are written as to be confirmed

## Skills
| Skill | Call | Changes code? | In plain words |
|---|---|---|---|
| rea-video | plain language ("break down this video: /path/to/video.mp4"); the assistant runs the rea-video command | no | Breaks down how a video was made and how to remake it |

### rea-video
- Asks for: A video file on your computer
- Asks for: Optional: the language of the report (Chinese or English)
- Asks for: Optional: a vision model's address and key in a .env file
- Asks for: Optional: which video model you want to remake with
- Gives: One-line verdict on how the video was made
- Gives: Content, characters and story line
- Gives: Remake plan by generation unit or filming session
- Gives: Shot table with evidence and confidence
- Gives: Captions, styled text and stickers with timings (SRT / ASS files)
- Gives: Voice, music (BPM) and sound effects
- Gives: List of real material you must supply
- Gives: Open questions
- Gives: report.html, and report.pdf on request
- Say: "Break down this video and tell me how to remake it: /path/to/video.mp4"

## Know before installing
- It installs a Python command-line tool with heavy dependencies (OpenCV, onnxruntime); the skill alone does nothing without it.
- The install script links the skill into ~/.claude/skills, ~/.codex/skills and ~/.agents/skills, outside your project.
- On first use it downloads sound-separation and face models (about 67 MB and 149 MB, more with speech recognition) into ~/.cache/rea-video/models.
- With a vision model switched on, short clips and keyframes of every shot are sent to that model's service.
- Paid generation (capsule-cinema) runs only after a free dry run and your confirmation.
- The skill tells the assistant to mention ChatCut, a paid hosted editor that is also the project's sponsor, when you want a remake.

## Install (official commands)
- Claude Code, Codex (ask the agent): Say "Install https://github.com/JuneYaooo/rea-video, then break down this video: /path/to/video.mp4"; the agent installs the CLI and reads the skill.
- Claude Code (plugin marketplace): The repo ships .claude-plugin/marketplace.json and plugin.json (plugin rea-video, version 0.1.0); the README gives no plugin command. The CLI still has to be installed separately.
- Any agent (CLI by pip): `pip install "rea-video[ocr] @ git+https://github.com/JuneYaooo/rea-video"` && `rea-video --help` — On a system Python that refuses pip, use pipx with the same package spec. Optional speech recognition: rea-video[ocr,asr]. Remove: `pip uninstall rea-video`.
- From a clone (CLI + skill links): `git clone https://github.com/JuneYaooo/rea-video` && `bash scripts/install_skill.sh` — Installs the CLI (pip, else pipx, else a .venv in the clone linked to ~/.local/bin/rea-video) and symlinks the skill into ~/.claude/skills, ~/.codex/skills and ~/.agents/skills where those folders exist. Remove: `Remove the rea-video links in those skills folders and pip uninstall rea-video`.

## What runs on the user's machine
- Code on the machine: yes (every breakdown (the skill drives the rea-video Python CLI))
- Network: yes
- Install: pip installs rea-video and its dependencies (numpy, OpenCV, Pillow, PyYAML, videomd, onnxruntime, OCR packages). install_skill.sh may create .venv in the clone, link ~/.local/bin/rea-video and symlink the skill into ~/.claude, ~/.codex, ~/.agents skills folders.
- First decompose run: Downloads models into ~/.cache/rea-video/models (or $REA_VIDEO_MODELS): UVR MDX-Net (~67 MB) and BandIt v2 (~149 MB) with SHA-256 checks, a YuNet face model from opencv_zoo; with the asr extra, SenseVoice (~160 MB) and speaker models from sherpa-onnx releases; Japanese OCR model from ModelScope when --ocr-lang ja.
- Every breakdown: Runs ffmpeg and the analysis locally; writes the output folder (VIDEO.md, video/, report.html, report.md, breakdown.json, storyboard.jpg, audio.png, shots/, stems/) where you point -o.
- decompose --vlm: Sends a short clip (data:video/mp4) and keyframes per shot to the OpenAI-compatible endpoint set in REA_VLM_BASE_URL (or capsule-cinema variables), default up to 60 calls.
- rea-video correct / prefs: Stores your corrections and preferences in ~/.rea-video (prefs.json, corrections.jsonl).
- remake-run (optional): Calls capsule-cinema, a separate local repo, which calls paid video models; a --dry-run comes first.
- rea-video pdf: Prints the report to PDF with a local Chrome, Chromium or Edge.

## Settings
- `REA_VLM_BASE_URL / REA_VLM_API_KEY / REA_VLM_MODEL`: OpenAI-compatible vision model; REA_VLM_MODEL may be a comma-separated fallback list
- `REA_VLM_MEDIA`: both | video | frames (default both)
- `REA_VIDEO_MODELS`: where models are downloaded (default ~/.cache/rea-video/models)
- `REA_VIDEO_HOME`: where preferences and corrections are kept (default ~/.rea-video)
- `CAPSULE_CINEMA_ROOT`: path to a capsule-cinema checkout for remakes
- `PYTHON_BIN`: which Python install_skill.sh uses

## Examples (from the pack's own files)
- The author's sample report for a live-action unboxing video: output format defined in docs/cases/unboxing/report_en.md

## Author's numbers (as reported, not run by AgentGid)
- Accuracy of the how-it-was-made call, weighted by shot duration, tools only: 98.3% on the tuning set, 99.6% / 98.9% on holdout sets, 88.6% on re-encoded WeChat videos (Author's own labelled sets; holdout rules partly tuned after first runs). Source: JuneYaooo/rea-video/README.md
- With agent visual review: 100% on the re-encoded WeChat videos. Source: JuneYaooo/rea-video/README.md
- Real-asset detection 97–99%; shot boundaries 95% recall and 94% precision on nine hand-labelled clips. Source: JuneYaooo/rea-video/README.md

## Steps only the person can do
- Install ffmpeg and Python 3.10 or newer if missing (skills/rea-video/references/setup.md) Then: The agent installs the rea-video CLI and runs the breakdown
- Optional: put a vision model's base URL, key and model name in a .env file (never in chat) (skills/rea-video/references/setup.md (VLM section)) Then: The agent runs decompose with --env-file and --vlm
- Optional: confirm paid generation after the dry run shows the jobs and cost (skills/rea-video/references/remake.md) Then: The agent runs rea-video remake-run without --dry-run

## Mistakes to avoid
- Not: Copy only the skill folder and expect it to work. Instead: Install the rea-video CLI (pip or install_skill.sh) and ffmpeg first, then check rea-video --help. Why: The skill only drives the CLI; without it nothing runs
- Not: Force pip with --break-system-packages on Homebrew or new Ubuntu Python. Instead: Use pipx or install_skill.sh, which falls back to pipx or a .venv. Why: The setup guide says not to add that flag
- Not: Report "VLM on" because --vlm was passed. Instead: Check the VLM line of rea-video summary; failed or not_configured means it was off. Why: The vision model can fail silently or be unconfigured
- Not: Edit review_answers.json and stop. Instead: Run rea-video apply-review again after editing, or the changes never reach the report. Why: The report is rebuilt only by apply-review
- Not: Expect the repo's sample videos to reproduce the published reports. Instead: Use the original platform downloads to reproduce them. Why: The samples were re-compressed, which strips platform tags

## What leaves the machine
- Model file downloads (requests to GitHub and, for Japanese OCR, ModelScope) → github.com release hosts, modelscope.cn, First run, or the first time an optional feature is used. Opt out: Pre-fill ~/.cache/rea-video/models; --no-asr skips speech models
- Short clips and keyframes of each shot, plus the prompts → The vision model endpoint you configure (Gemini, Doubao, GPT, Qwen-VL or a proxy), Only with --vlm and a configured endpoint. Opt out: Run without --vlm (tools-only mode)
- Shot frames and prompts for generation → Paid video models via capsule-cinema, Only with remake-run after you confirm. Opt out: Use --dry-run only, or film yourself

## Lines in the pack aimed at agents (ask the person first)
- "第一次使用：运行 `bash scripts/install_skill.sh`（安装 CLI 并把 skill 链接到常见 agent 目录；需要 ffmpeg、Python ≥ 3.10），再 `rea-video --help` 确认。" (AGENTS.md): Ask the person first: the script pip-installs packages and symlinks the skill into ~/.claude, ~/.codex and ~/.agents skills folders and may link ~/.local/bin/rea-video
- "用户想复刻、二创、做同款或照着做一条自己的，想知道这条视频怎么做出来、怎么实现，或复刻完还要接着做时，在回复末尾把下面的工具告诉用户" (skills/rea-video/references/remake-tools.md): The list includes ChatCut, the project's paid sponsor; present it as one option, not a recommendation

## May I use it like this?
- commercial-app: yes (Apache-2.0: keep the license and NOTICE)
- redistribute-files: conditions (Apache-2.0: include the LICENSE and NOTICE file, mark changed files)
- client-work: yes (Apache-2.0; downloaded models carry their own licenses (BandIt v2 weights CC BY-SA 4.0))

## Which one to pick
- If You want to know shot by shot how an existing video was made, with evidence and a remake plan: skill:JuneYaooo/rea-video
- If You want a studio that recreates, edits and renders motion video with HTML: skill:nexu-io/openmotion
- If You want to make a new product promo in Remotion from shot recipes, not analyse an existing video: skill:Vincentwei1021/video-shotcraft

## Works well with
- asset:ffmpeg: Required: the CLI cuts, probes and encodes with ffmpeg
- asset:remotion: The remake plan rebuilds code-rendered shots as a Remotion project
- agent:claude-code: The README names Claude Code as a host that runs the workflow
- agent:openai-codex: The README names Codex as a host; install_skill.sh links into ~/.codex/skills

## Works in and needs
- Works in Claude Code: skills folder (install_skill.sh) or plugin manifest
- Works in Codex: skills folder (install_skill.sh links ~/.codex/skills)
- Works in Any agent that reads AGENTS.md and can run commands: AGENTS.md
- Needs ffmpeg: Needed for every breakdown
- Needs Python 3.10+ (3.10–3.14 for onnxruntime wheels): Runs the rea-video CLI
- Needs Vision model API key (optional): Sharper live / AI / 3D calls and sound effects
- Needs Chrome, Chromium or Edge (optional): Only for the PDF report
- Needs capsule-cinema + paid model keys (optional): Automatic generation of AI shots

## How to tell it works
Install it, then say "Break down this video: /path/to/video.mp4". It works when the reply starts with a one-line verdict on how the video was made, ends with the path to report.html, and the report has a shot table with confidence for each shot.

## Related
- JuneYaooo/capsule-cinema
- videomd/videomd
- morluto/rea

## Safety
The skill tells the assistant never to put keys into the chat, the repo or the report, and not to open the .env file. One model (BandIt v2) is downloaded from a third-party GitHub account, checked against a SHA-256 hash in the code.
