Peter Yang’s AI Short-Video Workflow Relies on Review and Iteration

要約

30 秒でわかる

Peter Yang demonstrates a free video-short skill that turns a raw recording into an edit with captions, B-roll and motion graphics. The agent first proposes a scene-by-scene plan, including any paid generated assets, and the creator reviews its sources and costs. Yang then comments on framing and layout in a review page before approving a final render. He emphasizes that the example required several rounds of feedback rather than working in one shot. The video is sponsored by Gemini.

1The workflow has four stages, and generated assets may cost moneyYang installs the skill, supplies a raw recording, gives feedback and approves the render. Optional image, video and music generation requires an API key with billing.
2The agent proposes a scene plan and identifies sources before editingYang checks suggested B-roll and the places where the plan calls for paid generation. He adds a reference and a requested graphic.
3A review page turns framing problems into specific notesYang comments where his head is cropped and asks for a simpler layout. The agent makes revisions before he approves a render.
4The result takes multiple feedback roundsYang explicitly says the example was not a one-shot edit. The review interface makes iteration easier, but the creator still checks it.
5The skill includes a checklist for the opening and visualsYang says the first seconds should capture attention and visuals should add something to the words. He also describes an evaluation file for checking the output.

ポイント

Review the proposed scenes, B-roll sources and paid-generation plan before making the edit.
Give scene-specific feedback, then check the revised video before final rendering.
For longer recordings, specify how to handle pauses, retakes and filler without creating excessive jump cuts.
Yang recommends opening a local environment file for an API key instead of pasting it into chat.
Use a checklist to review whether the opening and visuals serve the story.

トピックGenerative media 画像と動画 Workflow automation

全文

チャプターごとに整えた全文を読む

元の文字起こしと対照

文字起こしのみ

Installing the short-video workflow

Yang introduces his free skill and optional paid generation. He demonstrates setting up dependencies and keeping an API key out of chat.

I demonstrate a free video-short skill that takes raw footage and helps create a short with captions, B-roll, motion graphics and optional generated video. My workflow is to install the skill, provide the recording, review and iterate, then render the result. This video is sponsored by Gemini.

I install the skill from its repository in an agent environment. On first use it checks for tools including FFmpeg, yt-dlp and HyperFrames. An optional Gemini API key enables generated images, video and music; those API calls require billing and can cost money.

For the API key, I ask the assistant to open a local environment file and enter the key there rather than pasting it into the chat. The key is optional if I do not want the generated media features.

From raw footage to a scene plan

The agent proposes cuts and visual sources. Yang reviews the plan and explains his prompt for cleaning longer recordings.

I provide a 54-second raw recording and ask the skill to make a short. It proposes a scene-by-scene plan, including full-screen speaker footage, B-roll and split-screen graphics. I can review the suggested sources and where the plan intends to generate paid media.

スライド 1 / 2The scene-by-scene edit planThe agent proposes a layout and visual for each part of the recording.

For longer recordings, I use a separate instruction to shorten pauses, remove repeats, false starts and filler, and avoid excessive choppy jump cuts. A basic version is included in the skill for cleaning short clips too.

I give the agent feedback on the plan, including a reference for a collage and a large number overlay. It produces a first cut. The important next step is not assuming the edit is finished but reviewing it scene by scene.

Reviewing and revising the draft

Yang gives scene-specific feedback on cropping and layout. He notes that the final result requires several rounds.

The skill opens a review page where I play the draft and leave comments at specific scenes. In two places my head is cropped, so I ask for a wider framing. I also ask to simplify part of the review-page layout.

After the agent revises the draft, I check that the framing and layout are fixed, then approve a final render. The version shown took several feedback rounds to reach; it was not produced in one shot.

スライド 2 / 2Reviewing the revised shortThe review page shows the clip, scene timeline and a place for feedback before final rendering.

What the skill adds and how it checks the result

The workflow can combine footage, captions, graphics and generated media. Yang describes a checklist for hooks and useful visuals.

The workflow can combine main speaker footage, tighter crops, split-screen B-roll, full-screen B-roll, captions and HTML motion graphics. Optional AI-generated video and music use the configured API service.

The skill file also includes editing practices: capture attention early, use visuals that add to the words and prefer footage that is legible on a phone. An evaluation file asks the agent to check the hook and other aspects before showing the result.

I am not a professional short-video editor, and others may make better shorts. My aim is to make editing more accessible so creators can spend more time on storytelling and producing content.