Production guides

Reverse any video: shots, prompts and a remake plan

“Reverse any video” is doing the rounds again, and it no longer means only playing a clip backwards. Since AI video tooling adopted the phrase — HeyGen’s HyperFrames editor ships a skill literally named “reverse-any-video” — it increasingly means rebuilding a finished video into its working parts: shots, production methods, prompts and a remake plan. This guide shows what a real reverse of any video looks like, how to run one locally for free, and what no tool can honestly promise.

Updated

Case evidence: Read the source breakdown report · View case frames and shot timings

Based on a published tool report. Production classifications are analytical judgments; suggested steps are not the original project or a verified recreation.

What “reverse any video” means now

Search the phrase and you meet two different jobs. The older one plays a clip backwards — most editors and mobile apps expose a one-click reverse, and a one-line FFmpeg command does it locally (fine for short clips; audio needs a separate areverse filter). If that is your goal, any of those will do. The newer meaning — the reason the phrase is spreading — reverses the video into how it was made: AI tooling now rebuilds finished clips into editable pieces and prompts, HeyGen’s open-source HyperFrames being the best-known example with its “reverse any video” skill. This guide is about that second job: producing a shot-by-shot account of any video — live action, AI, 3D, animation or code — with evidence you can check, plus the prompts and plan to remake it.

What reversing a video actually produces

Not a paragraph of prompt text. A real reverse of a video is a report: a shot table with start times, durations and representative frames; a production class for every shot — live action, AI video, digital human, 3D, 2D animation, stop motion, code rendering, screen recording, still-image push or freeze frame; the caption, sticker and picture-in-picture layers above the picture; and a sound map separating voice (human or TTS), music and effects. Each judgment carries its evidence and a confidence estimate, so you can argue with any row. At the end sit per-shot prompts and a remake plan that names the assets only you can supply.

The workflow, start to finish

rea-video is a free, open-source skill and CLI that runs inside an AI coding assistant such as Claude Code or Codex (Python 3.10+, FFmpeg). Point it at a local file: one command decomposes the video into the report above; a summary flags low-confidence shots for visual review; your judgments merge back into the report; and the remake plan groups shots into filming, generation and code work. Because it runs on your machine, unreleased material never leaves it — whether frames or audio go to a cloud model depends on the assistant you configure. This website shows finished examples and the guides; it does not accept uploads.

Why “any” video: because it judges per shot

“Any” is honest only if the analysis survives mixed production, and most real videos mix methods. In our music-video case (118 shots), the final verdict reads 53% AI video, 32% AI talking avatar, 13% still images with pan/zoom and 3% code rendering — one video, four production routes, each shot judged on its own evidence. The product-promo case sits at the other extreme: all nine shots read as code-rendered UI, so the remake route is a code rebuild. Live action, AI, 3D, stop motion and screen recordings each leave different evidence; per-shot classification is what lets one workflow cover them all.

What reversing cannot recover

Compressed video does not store the original prompts, project files or music licenses, and no tool can read them back out. A breakdown reconstructs production decisions from observable evidence, with uncertainty stated — shot #6 of the product-promo case stays labeled code rendering at 55% confidence and is listed as an open question. Treat every percentage as “what to verify next”, never as proof. And analysis is not permission: remaking with someone else’s footage, music or likeness still requires the rights.

Where this sits among the new tools

The “reverse video” space splits by job. Play-backwards tools and apps answer the old meaning. Upload-based video-to-prompt sites return prompt text quickly, but your file leaves your device and depth varies. HyperFrames’ reverse-any-video skill rebuilds a clip into an editable project inside its own editor, as a Pro feature. Chat models such as Gemini can describe a clip you show them — good for single questions. rea-video takes the report route: free, open source, local, evidence per shot, prompts and a remake plan you can hand to any tool. Pick by job: a quick answer, upload; rebuilding inside your editor ecosystem, use HyperFrames; a full account of how the video was made, run a breakdown.

Pertanyaan yang sering diajukan

What does “reverse any video” mean?

Two things. The classic meaning plays a clip backwards — a one-click feature in most editors. The newer meaning, spreading with AI video tooling in 2026, reverse-engineers how the video was made: shots, production methods, prompts and a remake plan. This site and the open-source rea-video tool are about the second meaning.

Is rea-video related to HeyGen’s reverse-any-video skill?

No. They are independent projects that share the phrase. HeyGen’s HyperFrames ships a Pro skill named “reverse-any-video” that rebuilds a video into an editable HyperFrames project. rea-video is an independent open-source project that produces an evidence-backed breakdown report — shot table, prompts and remake plan — and runs locally for free.

Can I reverse a video online for free?

For playing backwards, many free online tools exist. For reverse-engineering, upload-based prompt generators exist but send your file to someone else’s server. The open-source rea-video runs on your own machine for free; this website itself does not accept uploads.

Does it really work on any video?

Any local file FFmpeg can read — live action, AI-generated, 3D, animation, screen recordings. Mixed production is the norm and is handled per shot. Platform re-encoding can erase metadata, so judgments rely on visible and audible evidence with confidence estimates, not metadata.

A reusable analysis template

Goal check: play backwards, or understand production?
If backwards: editor one-click reverse / ffmpeg -vf reverse (short clips)
Input: local file, duration, shot count
Per shot: class + evidence + confidence
Layers: captions / stickers / audio (voice, music, sfx)
Prompts: one per generation unit
Remake plan: film / generate / code + required assets
Unrecovered: verbatim prompts, project file, music license

Practice with a real report

See “any” in practice: the music-video case mixes four production classes across 118 shots. Read its verdict, then pick one AI shot and one still-image shot and compare their evidence rows.

Open the case and download its report →

Continue your workflow