Kapwing Shows Claude Code Editing a 70-Second YouTube Short on Its Own
Kapwing says Claude Code, paired with a free custom skill called /shorts-editor, produced a finished 70-second YouTube Short in about 40 minutes. According to the company, the result had 98 clips across 13 tracks, and no manual edits were made.
- Kapwing says Claude Code with its free /shorts-editor skill produced a 70-second Short in about 40 minutes.
- The result had 98 clips across 13 tracks, with no manual edits, according to Kapwing.
- The skill uses FFmpeg, Whisper and HyperFrames, and learns a creator's style from 3 to 10 reference videos.
- Claude checks frames and audio of its own render and fixed issues such as loud music and overlapping text.
- Kapwing still advises human review before publishing.
What Kapwing tested
Kapwing set out to see how much of a Short's editing process Claude Code could handle without a human editor. The company points out that the job involves more than trimming: cutting pauses without clipping dialogue, matching screen recordings to narration, adding captions and graphics, balancing music and checking the final export. Even a minute-long video can need dozens of individual edits, according to Kapwing.
The team gave Claude raw footage, brand assets, sound effects and examples of previously published Shorts. Kapwing says it did not make a single manual edit to the output.
Claude Code does not work like Premiere Pro or Kapwing's own editor, Kapwing explains. It does not open a timeline and drag clips around. It uses commands and editing tools to analyze media and build the video programmatically.
How the /shorts-editor skill works
The skill combines several tools. Claude Code interprets instructions, studies reference videos and plans the edit. FFmpeg handles media analysis, cutting and audio processing. Whisper transcribes speech and helps find precise cut points. HyperFrames builds the timeline and renders the finished edit.
Kapwing describes a five-step process: install the skill, organize media, give Claude examples of your editing style, run the edit, and let it inspect the result. The skill is a 4 KB ZIP file, and Claude Code must already be installed. Users also need FFmpeg, Whisper, Node.js 18 or newer, and HyperFrames. The skill checks for these dependencies on startup.
Media goes into a project folder with subfolders for footage, B-roll, music, sound effects, brand assets and references. Kapwing recommends one to three licensed music tracks and three to ten finished videos as style references. Separate screen recordings for each step or topic make it easier for Claude to match clips to narration.
The command /shorts-editor learn my style from ./references makes Claude analyze the reference videos using metadata, extracted frames, timing information and audio measurements. It writes a reusable guide, references/layout-guide.md, describing layouts, when elements appear, how long they stay and which transitions or sound effects go with them.
Editing and self-checking
To edit, the user opens Claude Code in the project folder and enters /shorts-editor edit this folder. Claude first transcribes the talking-head footage with word-level timestamps. It uses them to find pauses and repeated takes, keeping complete sentences and the final take when a line was recorded more than once.
The skill assembles a HyperFrames project and renders at 1080 × 1920, the vertical format for YouTube Shorts and Instagram Reels. It saves three files: output/final.mp4, output/timeline.json with the clips and tracks, and output/edit-notes.md, a record of the layouts chosen and why.
Claude then checks its own work by extracting frames, comparing exported audio with the original dialogue and measuring audio levels. According to Kapwing, this caught several issues:
- A full-screen talking-head shot covered the intended layouts for 49 seconds. Claude found two elements sharing a name, renamed one and rendered again.
- Part of a wall lamp appeared in the presenter cutout and some hair was clipped. Claude cleaned the cutout and checked the affected frames.
- Text and labels overlapped the edges of graphic cards. Claude resized and repositioned the cards.
- Music was too loud during quieter speech. Claude adjusted the music loop and lowered it beneath the voice.
What it means for creators
The setup and style learning only need to happen once, Kapwing says, and the same configuration can be reused with new footage. That makes the approach most relevant to creators who publish Shorts in a consistent format.
Kapwing is clear about the limits: the workflow does not remove the need for human review. Creators should still watch the finished MP4 before publishing, especially when it contains brand names or product demonstrations. The results come from one company's own test of a single Short, so they show what is possible rather than a general benchmark.
