Production guides

How to reverse engineer video prompts from a reference

To reverse engineer video prompts, describe how each shot changes over time, then separate what belongs in generation from what belongs in editing. A finished clip is evidence for a new production brief, not a reliable copy of the original prompt. The dialogue report below shows why that distinction matters.

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.

Start with the actual timeline

The dual-role report covers a 19-second (0:19) vertical video with 10 shots. It classifies five character shots as AI-generated and the other five as code or template rendering built around interface screenshots, for an overall mix of 59% code/template rendering and 41% AI-generated video. Those are report judgments, not verified production history. For prompt writing, the important finding is the alternation: a character moment is followed by a screen or interface segment. One prompt for all 19 seconds would mix two different production tasks.

Choose one shot before writing a sequence

Shot 2 runs from 00:01.375 to 00:02.833 (about 1.5 seconds); shot 3 from 00:02.833 to 00:03.875 (about 1.0 second). The report describes a short-haired young woman holding a phone on a sofa, and a second character in a newsboy cap holding a white mug in an armchair. Make a separate brief for each: subject reference, pose, framing, action, background and duration. The report itself suggests generating about 4 seconds for shot 2 and trimming to 1.5, because very short shots are unstable when generated alone. Keep the observed timeline as the edit target, not as a duration the generator must hit.

Describe motion without inventing a camera setup

For the phone shot, the report's prompt has the woman look up from her phone toward frame right, eyebrows raised and mouth open, while the phone moves slightly, ending on her surprised face, with a locked-off camera. That instruction keeps subject motion separate from a fixed camera. Avoid adding an orbit, a push-in and a rapid pose change simply to make the prompt sound cinematic. The report's lighting notes also disagree: the pixel measurement reads the light as cool, while the visual review describes a warm floor lamp mixed with cool phone-screen light. Inspect the reference frame and record what is visible rather than naming a specific lamp or lens as fact.

Keep UI, captions and voice outside the picture prompt

The report groups character shots 2, 3, 5, 7 and 9 separately from interface shots 1, 4, 6, 8 and 10. Its AI plan turns the five character shots into four generation units: shots 5 and 7, which share the living-room set, form one multi-shot unit, while shots 2, 3 and 9 are generated one at a time. Use that division to plan your work: prepare character references, capture your own interface, and add readable labels in an editor or code composition. On sound, the report leans toward the dialogue having been generated with the picture (65%, unconfirmed) and found no music bed; whichever route you take, keep the lines in a separate timed list so you can adjust pauses without rewriting every visual description. The grouping is a planning aid, not a promise that every model supports multi-shot generation.

Test consistency, then restore the edit

Start with one character shot and check face consistency, hand-object contact, framing and background stability. If the result fails, change the instruction that controls the failure instead of appending unrelated adjectives. Once the character material works, place the UI inserts back between the shots and align speech with the revised cut. This article explains how to use the report; it does not claim a tested recreation or recover the original model settings.

A reusable analysis template

Shot ID and reference time
Character reference / pose / object interaction
Framing / camera motion / subject motion
Visible lighting and background
Target duration / start state / end state
Separate tracks: dialogue / UI / captions
Test result / uncertainty / next adjustment

Practice with a real report

Read the original generation-unit and code-compositing sections alongside the shot table. They show how the same timeline can require different instructions for characters and UI.

Open the case and download its report →

Continue your workflow