← Back to Blog
Tutorial

Reverse-Engineering a Viral AI Video, Shot by Shot: A Real Walkthrough

November 13, 2026·9 min read·By amine.smpc

Every few days a short AI video goes viral and the comments fill with the same question: "what prompts did you use?" The honest answer is usually dozens of them, one per shot. This is a walkthrough of what the Video Prompt Extractor recovers from a real video, with the actual numbers from our run, and the two things we learned to check every time.

The video we used

A 42-second vertical (9:16) comedy skit: two toddlers in streetwear sit in lawn chairs in a backyard, decide to dig a pool to impress girls, strike oil instead, and celebrate. It is a good test because it has everything that makes extraction hard: fast cuts, two recurring characters, dialogue going back and forth, and a change of location halfway through.

Choosing the options

You can upload the file or paste a link. Before running, a row of options sets what you get back. We left most of them alone; these are the ones that matter:

  • Shots: Auto lets the engine follow the real cuts. Pick a number only if you want to merge or split.
  • Style: "Match the video" describes what is on screen. Choose Cinematic, Anime or another style to restyle every prompt instead.
  • Frame: rewrites the prompts for another aspect ratio, useful when turning a vertical video into a 16:9 one.
  • Part: analyse one section of a long video instead of all of it.
  • Brief: a sentence of creative direction, such as "same story, set in Tokyo at night". It changes what the prompts describe while keeping the shots, timing and camera work.

What came back: the shots

The run took under a minute. It found 28 shots, lined up with the real cuts in the video, and each shot came back with:

  • a frame from the shot and its exact start and end time;
  • the camera: shot size, lens and movement, for example "close-up · 50mm, shallow depth of field · slight handheld push-in";
  • who is in it, by character name;
  • an image prompt for the still and a motion prompt for the clip;
  • the dialogue, attributed to the character who says it, and the sound.

The split into image and motion prompts follows how most people work today: generate a still you are happy with, then animate it. The first shot's motion prompt was a single line — a handheld push-in on the boy finishing a sip from a tiny bottle, outdoor daylight — which is about the right length for a 1–2 second shot.

Character sheets and the master prompt

Alongside the shots came a character design sheet for each of the two boys: a turnaround prompt (front, three-quarter, side, back) describing age, face, hair and clothes, written to produce a clean reference image. Each sheet also lists the shots the character appears in. These sheets are what keep the faces consistent when you remake the video; in Google Flow they become the portraits of saved Characters.

Finally a master prompt describes the whole video in one paragraph, and again as structured JSON for tools that take fields. It is the quickest way to make "a video like this one" without going shot by shot. The voiceover and dialogue also come back as one script you can copy.

What to check before you use it

Two things are worth thirty seconds every time.

1. Who is who

In one of our early runs the model gave the two boys' names the wrong way round: the line "Yo Benny, this stuff hitting" was given to Benny himself. Names drive everything downstream — the character sheets and which reference goes into which shot — so one swap mislabels the whole video. A quick test: nobody calls themselves by their own name, so a line addressed to its own speaker means the names are swapped. Read the cast against the frames before you generate.

2. The notes

When the engine merges or repairs anything — shots too short to keep, timings that overlapped — it says so in a Notes list rather than silently returning fewer shots. If a note is there, read it.

Anything you want to change can be asked for in plain words: "make every shot golden hour", "split shot 2 into a wide and a close-up". For the first hour after extracting, a change that needs another look at the video can use it.

Remaking it

  1. Generate a reference image from each character sheet and save it as a Character in Flow, named exactly as in the extraction.
  2. Generate the stills from the image prompts, attaching the right characters to each shot.
  3. Animate each still with its motion prompt; keep clip lengths close to the original shot lengths.
  4. Cut them together on the timings from the shot list. The cut-list download (EDL) opens in Premiere or DaVinci.

Doing twenty-eight shots by hand is an afternoon. With AutoFlow Studio the extraction opens as a canvas with every shot's prompts already wired in, and AutoFlow queues the whole batch on Flow. Either way, the point of reverse-engineering is not to copy someone's video frame for frame, but to learn its structure — the pacing, the shot choices, the way the joke lands — and use it for your own story.

Want the "why" instead of the "how"? Why it works runs the same video through a different question: the hook, the pacing measured from the cuts, and where viewers are likely to leave.

Related guides