WhoBuilt

How to Edit YouTube Shorts With AI

Oct 21, 2026· 14 min· 4.0K views·Watch on YouTube
#Agent Skills#Human-in-the-Loop Systems#Generative Media#Multimodal Models#AI Security

Speaker

Peter Yangproduct lead at OpenAI

Peter Yang, creator of open-source AI skills, demonstrates how to turn a raw recording into a YouTube Short using his free video short skill, GitHub tools, and optional Gemini services.

Install the editing skill

Yang installs the video short skill from github.com/petergyang/videoshort. It uses FFmpeg for video cutting and encoding, yt-dlp to find and download B-roll, and HyperFrames for transcription and motion graphics. Gemini API access is optional for generated images, video, and music, and Yang advises putting the key in an .env file instead of chat.

Generate a draft and review the plan

Yang uploads a 54-second talking-head clip and asks the skill to make a visually engaging Short. The AI proposes cuts and a scene-by-scene layout, then lists B-roll sources so he can review them. He also describes an editing prompt for longer videos that trims pauses, retakes, and filler words while avoiding choppy cuts.

Iterate with frame-by-frame feedback

A review page lets Yang watch the draft and leave comments on specific frames. He asks the AI to zoom out where his head is cropped, add a 200-billion-views overlay to a Shorts collage, and simplify the review page. After further edits, he approves the final render, while noting that several rounds of feedback were needed.

Combine footage, graphics, and checks

The skill can mix A-roll, tighter punch-ins, split-screen footage, and full-screen B-roll, with captions and HTML motion graphics. Online B-roll is available alongside optional Gemini-generated video and Lyria music. Yang points to guidance in the skill file, including capturing attention in the first three seconds, and an eval file that prompts the AI to check its own work.

Takeaways

  • Build review into AI editing: frame-specific comments make visual corrections easier to communicate.
  • Treat generated edits as drafts; expect multiple feedback rounds before rendering a final Short.
  • Keep API keys in environment files instead of pasting them into AI chat.
  • Design mobile visuals around the spoken message, with the strongest visual in the first three seconds.
  • Use evaluation criteria such as hook strength and whether each visual adds information beyond the words.

Mentioned in this talk

Google XYouTubeBehind the CraftVideo Short SkillGeminiAntigravity agentVeoGemini APIHyperFramesGoogle AI StudioLyriaNo AI Slopyt-dlpFFmpeghuman reviewNate Hürk

Search inside this talk

What was claimed 14

Related talks