I taught Claude how to edit videos and it's INSANE

Video Flow

1. 0100:00

   A six-job system replaces one-shot editing

   Claude handled captions, motion graphics, and cards without the creator touching an editing timeline, but the result was not perfect and still needed human notes. Splitting the work into six jobs and saving the instructions as a skill helped avoid the generic look of earlier AI edits.

   Claude Code
2. 0201:50

   Research hooks and write the script

   Claude studies successful videos in the creator’s niche, extracts their hooks and formats into a reusable library, and uses those patterns to write scripts. The creator then reads the script to camera.

   Claude Code
3. 0302:19

   Transcribe and cut raw footage

   Claude uses Whisper to transcribe raw footage with word-level timestamps, then keeps the last clean take of each line and uses FFmpeg to remove the rest. A short test recording with deliberate mistakes can check whether the retakes are removed and listed.

   WhisperFFmpegClaude Code
4. 0403:03

   Set the visual direction and map each idea

   Claude can derive design guidelines or mimic visual references from screenshots, screen recordings, or video files; PromptMotion offers examples with their prompts. Before building graphics, Claude creates a beat map linking each script idea to what appears onscreen, with one readable graphic per idea rather than one per sentence.

   Claude CodePromptMotion
5. 0504:17

   Build graphics in parallel and control renders

   Claude splits a long video into sections and assigns each to a sub-agent, which builds graphics with code using Hyperframes to turn web pages into video and create word-by-word captions. Planning and creative direction go to more capable, costlier models such as Opus and Fable, while faster, cheaper Sonnet models handle building; limiting renders to four at a time, using half-size drafts, and waiting for approval before full-quality renders avoids overloading the computer.

   HyperframesClaude CodeOpusFable+1
6. 0605:45

   Review the cut against a mistake checklist

   Claude checks ten frames per second for six problems: repeated animation, brief flashes, late-arriving text, graphics over the creator’s face, duplicated spoken words, and lip-sync drift. A checklist based on the creator’s own recurring corrections makes the review more person-like.

   Claude Code
7. 0706:16

   Turn feedback into better instructions

   After each cut, Claude applies the creator’s notes and identifies which instruction would have prevented each mistake, then rewrites the skill so future videos are less likely to repeat it. The creator still films and makes the final call; the accumulated notes, rather than a single prompt, shaped the result.

   Claude Code
8. 0807:02

   Cost, time, and the limits of the result

   This edit used about 600 million tokens—roughly 100 million for Opus planning and design direction and 500 million for Sonnet building—and took about six hours of Claude work, excluding time spent watching cuts and giving notes. Reusing graphics, tools, and the skill reduced usage from the first AI edit’s roughly 1.4 billion tokens and 12 hours; the creator estimates the current edit at about $160 at API prices, while their Claude Max 20X subscription covered it.

   Claude Max 20X