← 返回内容列表
𝕏X采集原文

SageCut筛选视频帧供Opus审查

10-26 00:40 · 北京时间

AI 编程

有实践材料可继续查看

教程线索

原文与观测指标

1 个来源
𝕏X

SageCut筛选视频帧供Opus审查

Shann³
查看原文
点赞149采集时的累计值
每小时增长待观察赞 / 小时 · 观测区间平均值
观测增量待观察两次有效记录之间的变化

发布时间:2026-10-26 00:40 · 北京时间

指标采集:10-26 17:35

简短概述

作者介绍Claude Code插件SageCut:先用代码检查Remotion视频规格,再由Sage逐帧判断,仅将选中帧交给Opus 5.5复核。作者称此法可降低审查成本与耗时,并支持品牌规范及自定义检查项。

AI 概述,仅依据已采集材料,未经独立核验。
采集原文 展开 / 收起

how to improve your Opus 5.5 motion videos and make the review much cheaper if you use Opus 5.5 to review every frame of the videos you make, it...

how to improve your Opus 5.5 motion videos and make the review much cheaper if you use Opus 5.5 to review every frame of the videos you make, it quickly gets expensive and slow... so add a fast, cheap decision model to your workflow, and let it pick the frames Opus needs to see here's the playbook, save it: 1. add the decision model to your video workflow I use SageCut, a Claude Code plugin that runs the Sage decision model on Remotion videos install it in Claude Code > /plugin marketplace add levantolabs/sagecut-release > /plugin install sagecut@levantolabs 2. open a new project folder, or your current agent chat, and prompt: "Use SageCut to review my Remotion video at out/video .mp4" 3. set your own requirements the first time, it asks for your brand (your website, a brand kit folder or 5 questions), and it will then be able to review every video against your colors, fonts, motion style and the words to avoid you can also add your own yes/no questions for Sage to the review criteria. I want no slop in my videos, so I added my anti-visual-slop .md rule file to SageCut review skill what happens on each video: > SageCut measures every frame in code: text size, blank frames, frozen frames, off-brand colors and the delivery spec > Sage reads every frame and answers yes/no questions about it, including yours > only the frames Sage selects go to the Opus 5.5 critics > at the end, it shows what the review cost and what it saved on Levanto's reference build, an Opus 5.5 critic cost about $0.10 and 8.6 seconds for each frame it viewed (repo in comments)