AI column

Video@donghyeok99Staff

rea-video: a skill that tells how each shot was made

Given an AI remake of a TikTok ad, it answered: all AI, and its 8 shots take 3 generations

A playhead runs over the film, cuts it into shots tagged filmed, AI and code, splits voice from music, and writes the mix into report.html.
A playhead runs over the film, cuts it into shots tagged filmed, AI and code, splits voice from music, and writes the mix into report.html.
Who
Open source by JuneYaooo (Apache-2.0), for creators and marketers who see a good video and want to know how it was made.
When
Published on 29 October 2026; no tagged release yet, with commits landing daily (as of 7 November).
Where
Runs in Claude Code, Codex or any agent that runs commands, on your own machine with ffmpeg and Python 3.10+. A vision model is optional.

What was hard

A good short ad makes you want to try one yourself, but you can't tell by eye which parts were filmed and which were generated. Downloads from the platform are recompressed, so the clues are gone, and caption timing or the music's tempo had to be measured by hand.

What it does about it

Give it a video and it cuts it into shots, separates voice, music and effects, and labels each shot as one of 11 kinds, from live footage and AI video to 3D and code renders, with the evidence. Measurements come first: frame cadence, sensor noise and AIGC tags in the metadata outrank how a shot looks. Shots where the signals disagree go to your agent to look at, and the result is a remake plan split into what to film, what to generate and what to code.

Left: a TikTok ad shot on a phone. Right: an AI remake with another product.
Left: a TikTok ad shot on a phone. Right: an AI remake with another product.
After the breakdown, a remake: original and remakes playing in sync.
After the breakdown, a remake: original and remakes playing in sync.
A 19-second clip mixing AI characters, code UI and real screen capture.
A 19-second clip mixing AI characters, code UI and real screen capture.

What people made with it

GitHub@wusanjay2026-11-05

Ran a 123-second Douyin talking-head video and had the agent review all 15 shots, found that dark close-ups of a real person stayed labelled AI, and sent a fix.

一条 123s 的抖音口播视频(暗背景、单侧灯条、近景、无运镜),agent 看图复核后对 15 个镜头逐镜给出 label_probs

How it runs inside

  1. 1

    You → Claude Code

    One line: the repo and a video pathWith the README's line, the agent installs the CLI and learns the workflow.README.md
  2. 2

    Claude Code → rea-video skill (SKILL.md)

    Reads the standard flowDecompose, summary, visual review, apply, and the rule to write 'unconfirmed' rather than guess.skills/rea-video/SKILL.md
  3. 3

    Claude Code → rea-video CLI

    rea-video decompose ref.mp4 -o out/refThe CLI takes the command and starts the pipeline.rea_video/cli.py
  4. 4

    rea-video CLI → breakdown.json

    Shot cuts, voice and music apartIt cuts at scene changes and separates the audio into voice, music and effects.rea_video/shots.py
  5. 5

    rea-video CLI → breakdown.json

    One of 11 kinds per shotHard evidence such as AIGC tags and frame cadence outranks 'looks like' when it scores each shot.rea_video/classify.py
  6. 6

    Claude Code → breakdown.json

    rea-video review, then apply-reviewFor low-confidence shots the agent looks at the frames itself and its answers are merged in.rea_video/review.py
  7. 7

    breakdown.json → report.html

    Evidence, prompts and a remake planThe report splits the remake into what to film, what to generate and what to code.rea_video/report.py

What goes in, what comes out

A TikTok ad shot on a phoneAll live footage, one set, two camera positions
Its AI remake with another productAll AI, 8 shots in 3 generations
A 70-second product promo83% code-driven motion graphics, one Remotion project
A 165-second AI music videoSong flagged as AI-generated (it was made with Suno)

Is it for you

Good for

  • Learning how a well-performing short ad was shot
  • Deciding with evidence whether footage is real or generated
  • Splitting a remake into filming, generation and code before you price it
  • Pulling caption timing out as SRT or ASS

Not for

  • The detailed docs are in Chinese for now; reports come out in Chinese or English (per the README).
  • On recompressed platform downloads the tools alone reach 88.6%, so add a vision model or agent review (per the README).
  • It doesn't render video; automatic generation is capsule-cinema's job and calls paid models.
  • It isn't for reposting other people's footage; remake with material you have the rights to.

Sign in to post, comment and like.Sign in

Report

Reason

Read the community rules