# rea-video: An Agent Skill That Tells You How Any Video Was Made, Shot by Shot > Source: > Published: 2026-11-06 00:00:00+00:00 A new GitHub repo is gaining stars fast: **rea-video**, an agent skill that takes any video and tells you, shot by shot, whether it was filmed, generated by AI, rendered in 3D or built in code. According to the [repository](https://github.com/JuneYaooo/rea-video?ref=explainx), it has about 2.2k stars and 318 forks, uses the Apache-2.0 license, and was built to run inside Claude Code, Codex or any agent that can run commands. Its trending count is 2,228 stars since October 29. It also reports how captions, cuts and audio were made, then writes a plan to remake the video cheaply. This post covers what it does, how the accuracy figures were produced, and where the author says it fails. It extends the idea behind REA (Reverse Engineer Anything), the skill we covered in our [install guide](https://explainx.ai/blog/how-to-install-use-rea-reverse-engineer-anything-skill-claude-code-guide-2026) and the [projects built with it](https://explainx.ai/blog/top-10-projects-built-with-rea-reverse-engineer-anything-skill-2026). The README credits REA as its inspiration: trace from what you see back to how it was made, back every claim with evidence, and verify by measurement. ## TL;DR: the questions people ask first | Question | Answer | |---|---| | What does it do? | Splits a video into shots and labels each as live, AI video, avatar, 3D, code or template, or screen recording | | What else does it report? | Captions with timings (SRT and ASS export), cuts, stickers, audio type, music BPM, sound-effect candidates | | What do I get back? | A one-line summary, a remake plan, and a full `report.html` with per-shot frames and evidence | | License | Apache-2.0 | | Needs | ffmpeg and Python 3.10 or newer | | Free? | Tools-only mode is offline and free; a video model like Gemini is optional | | Does it render the remake? | No. It plans; companion tools generate or you film and code | | Accuracy claim | 98.3 to 99.6 percent on the author's own sets (see caveats below) | ## What does rea-video actually break down? The README groups videos into types and says what you get for each. Talking-head and tabletop footage gets setups and camera positions, framing, camera moves, lighting (12 setups with gear lists) and color grade. AI video gets the evidence for that call, a reverse-engineered prompt per shot, how to batch shots into generations, and which model to use. Avatar clips are flagged with what portrait and audio you would need. 3D and animation clips get a verdict on whether to use a video model or render from Blender or Three.js code. Kinetic type, UI demos and slide-style cards are labeled as code or template and are meant to be rebuilt as a Remotion project. Screen recordings are flagged as real material you must record yourself. Film frames laid out as comic panels with one frame highlighted, showing how rea-video splits a video into shots Mixed videos are judged shot by shot and reassembled on the original timeline. The README example is a 19-second clip with 10 shots: AI characters talking, code-built UI and 8 seconds of real screen recording. Its five dialogue shots share one set, so a single generation covers them. Beyond the picture, every video also gets a text and edit layer. Captions come with start and end times, color, outline and entrance style. Stickers, emoji and picture-in-picture are placed on the timeline. Cuts are measured for speed and whether they land on the beat. ## How does it work? The README describes four steps. First it cuts the video into shots and separates audio into voice, music and effects. Then it gathers evidence: it measures frame cadence, sensor noise, whether the whole image pans as one plane, and AIGC tags in the file metadata. It also looks at the images for hands, text and physics, and listens for BPM, key, and singing versus speech. Third, it decides. Hard evidence such as AIGC tags and frame cadence outranks "looks like", and conflicts are flagged for the agent to review visually. Last, it plans the remake, batching shots by what the target video model can do, such as maximum clip length, multi-shot support and how many reference images it accepts. Capability tables for video, image and audio models live in the repo under `rea_video/capabilities/`. The design choice matters. A pure vision-model guess about "does this look AI" is easy to fool. Measuring frame repetition or noise is harder to fool on original files, which is why tools-only mode does well there. The same logic also explains where it breaks, covered below. ## How do I install and run it? The README is written for agents. In Claude Code or Codex, you say: ``` Install https://github.com/JuneYaooo/rea-video, then break down this video: /path/to/video.mp4 ``` To install by hand, the README gives these steps: ``` brew install ffmpeg # or: sudo apt install ffmpeg git clone https://github.com/JuneYaooo/rea-video && cd rea-video bash scripts/install_skill.sh rea-video --help ``` The install script sets up the CLI with pip, pipx or a local virtual environment if pip is blocked, and links the skill into the Claude, Codex and agents skill folders. Reports come out in the language you ask for, Chinese or English. The detailed docs are in Chinese for now, and your agent reads them and answers in your language. If you are new to this packaging, our [agent skills guide](https://explainx.ai/blog/what-are-agent-skills-complete-guide) explains how a skill folder loads into an agent. ## How accurate is it? Read the eval before the headline The headline table in the README measures "how it was made" accuracy, weighted by shot duration: | Method | Tuning set (15 videos) | Holdout (9) / Holdout 2 (7) | Re-encoded WeChat videos (10) | |---|---|---|---| | Tools only (offline, free) | 98.3% | 99.6% / 98.9% | 88.6% | | Plus a vision model (Gemini) | not given | 99.6% / 99.8% | 99.6% | | Plus agent visual review | not given | 99.6% / 99.8% | 100% | The author's own results file, `eval/RESULTS.md`, is more candid than the summary. It says the run used deterministic forensics and OCR with no vision model, re-decoded on 2026-11-02. The tuning set is where the rules were designed. The first holdout is no longer fully independent, because two first-round errors led to a frame-rate rule. The second holdout first scored only 75.6 percent, was fixed afterward, and the file also calls it not independent. The WeChat set is labeled a development set, because rules for platform transcoding were written after reviewing those videos. The file also lists a second metric on that set, a score that rose from 0.79 to 0.92 between versions 0.2 and 0.3. In short: the numbers are real measurements, but all of them come from the author's own clips (31 in the three main sets, 10 more in the WeChat set), and none is a third-party test. Two points from the results file are worth knowing: - **Recompression hurts.** Videos downloaded from platforms are re-encoded, which wipes out the noise evidence. The tools-only score on the 10 re-encoded videos is 88.6 percent. In one case, an unboxing video that was all real footage was called 35 percent AI by the offline tool, and agent review brought it to 97.4 percent accuracy on that clip. - **Audio is the weak layer.** The results file reports 2 of 4 correct on whether sound effects exist, and 2 of 5 on music source. It says signal features cannot separate sound effects from speech or music onsets. Music from Suno and library tracks is hard to tell apart by signal, so the report says "leans library, to be confirmed" rather than claiming certainty. These limits are exactly what you would expect from a detector built on measurable traces. They also show the author reporting misses, which is a good sign for a repo of this age. ## What can you do with it today? The README lists six example reports you can read without installing anything. They include a 70-second product promo that is 83 percent code-driven motion graphics and 12 percent screen recording, so no video model is needed and one Remotion project would do it. A 67-second 3D explainer is flagged as doable in 10 video generations, or by having an LLM write Blender or Three.js code. An AI music video of 165 seconds and 86 shots was correctly flagged as AI-made, and the README says it was created with Suno. The vision model called some shots live-action drama, but the agent spotted fake Chinese characters on a scroll and ruled them AI. A hybrid clip was 75 percent real and 25 percent AI, with the AI segment caught by a Jimeng watermark and garbled text on a gift box. Practical uses we can see from the README: 1. **Learn how a viral ad was made.** Run it on a clip and read the shot list, camera positions and lighting notes before you plan your own shoot. 2. **Price a remake before you start.** The split between "film", "generate" and "code" tells you whether you need a video model, a camera or a Remotion project. 3. **Check a suspicious clip.** The evidence list (AIGC metadata tags, frame cadence, watermarks, garbled text) is a useful starting point. It is not proof. For provenance from the generator's side, see our guide to[Google's SynthID verification site](https://explainx.ai/blog/google-synthid-website-public-ai-media-verification-guide-2026) . 4. **Prepare agent edits.** Pair the report with an agent video editor such as[Kino](https://explainx.ai/blog/kino-ai-video-editor-claude-code-codex-agents-shared-timeline-review-links-plugin-2026) or the[Lemo skill for short films](https://explainx.ai/blog/lemo-opuscar-claude-code-skill-short-films-no-video-model-43-styles-2026) . Microphone with sound waves, standing for the audio breakdown that rea-video produces for each video ## How does the remake side work? Scissors trimming a filmstrip, standing for the remake plan that rea-video builds from a shot list After a breakdown, the README points to a remake recipe in a separate project, cheap-claude-video. Camera, motion, pacing and transitions follow the original, and only the elements you choose are swapped. The author's demo results use a self-hosted MiniMax model, with the original shot as the motion reference and keyframes with swapped elements set the look. Three remakes are shown against versions in a WeChat article from OiiOii, once with the same swap and once with a fuller element swap, such as a fox replacing a clay girl, or a camping table replacing a kitchen. With the capsule-cinema tool, AI shots are generated and placed back on the timeline automatically. The README says it runs a free dry run first and only calls paid models after you confirm. Live shots you film yourself, and code shots the agent writes as a Remotion project, as described in the repo's `docs/replicate.md`. If you are choosing a video model for those AI shots, our posts on [Kandinsky 6 Video](https://explainx.ai/blog/kandinsky-6-video-open-source-29b-audio-video-mit-license-2026) show one open option with audio. ## How is it different from one-click remake tools? The README answers this itself. It says Hypit takes a reference video and makes a new one by reading frame grids, generating with the models it supports and rendering captions in HTML. According to the README, that approach does not ask how the original was made, so a live-action source goes through the same pipeline and it will not say how many cameras or what lighting were used. rea-video does the opposite: it renders nothing, and explains the original. The author's advice is plain: for a quick lookalike, a one-click tool is faster; to understand and remake cheaply, use rea-video; they work fine together. We have not tested Hypit, so treat this as the author's comparison. ## What are the risks and limits? - **Not a deepfake judge.** The author reports 97 to 99 percent on real-asset detection and says all shot boundaries were correct. But these are small sets, and a determined faker can strip metadata and add noise. Use the report as one signal. - **Platform recompression.** Expect lower confidence on downloads from social apps unless you add a vision model. - **Third-party model data.** If you add Gemini or Doubao, your video frames go to that provider under that provider's terms. Check them before you analyze client footage. - **Rights.** The README states that you must not reuse other people's footage, music, characters or likeness, and must follow platform and local AI-labeling rules. For the broader legal picture of reverse engineering with AI, read our[clean-room checklist](https://explainx.ai/blog/is-ai-reverse-engineering-legal-clean-room-checklist-rea-guide-2026) . - **Young project.** The repo shows 53 commits and docs mostly in Chinese. Pin a version before you build a workflow on it. ## Why it matters Two trends meet in this repo. Video is now a mix of live footage, generated clips and code-rendered graphics, often in one 20-second ad. Provenance has become a real question for platforms, courts and newsrooms: in October a [Nikon Small World in Motion winner was disqualified](https://explainx.ai/blog/nikon-small-world-in-motion-winner-disqualified-ai-generated-video-2026) for AI-generated video. Tools that measure traces, rather than trusting how a clip looks, will become common. For builders, the other lesson is the packaging. rea-video is a skill that gives an agent a workflow, measurement tools and reference tables, and lets the agent finish the judgment calls by looking at the frames. That is the same pattern as [leanrea](https://explainx.ai/blog/leanrea-8kb-skill-replaces-rea-mcp-38-percent-faster-35-percent-cheaper-claimed-vs-verified-2026), and a good template if you plan to ship your own skill. ## Related reading Official sources: the [rea-video repository](https://github.com/JuneYaooo/rea-video?ref=explainx) and its [evaluation results file](https://github.com/JuneYaooo/rea-video/blob/main/eval/RESULTS.md?ref=explainx). Hacker News has a [short thread](https://news.ycombinator.com/item?id=50368605&ref=explainx) with no comments yet. *Star counts, accuracy figures and repository details are accurate as of November 6, 2026 and may change as the project updates.*