claude-video：给代理看视频的 /watch 技能，下载、抽帧、转写后再交给 Claude 或 Codex
MIT 许可，Python，约 ★20188。把视频 URL 或本地文件变成带时间戳的帧和转写；优先用原片字幕，没有字幕可以用本地 WhisperX 或 Groq/OpenAI。可装进 Claude Code（桌面应用的 Code 会话、VS Code 或终端）、Codex 和其他 Agent Skills 宿主；作者写明它不能在 Claude Chat 或 Cowork 里用。抓取数据里的日均约 +78。适合需要让编码代理看演示录像、会议录屏或产品视频的团队在本地使用。
https://github.com/bradautomates/claude-videoReposted. Will be removed on request. All rights belong to the original author.
More from this batch
An independent tester added unreleased Qwen-4-Flash-Early (codename hawk) to GameFeel, with recapped scores of 49.9 vs 47.2 for Opus 5.5 XHigh
follistatindev says Alibaba granted early API access. Recaps of the tester’s charts list Flash-Early at 49.9, Opus 5.5 XHigh at 47.2 and an …
DoubleAI: an agent redesigns a network during training; DeepSeek-V4.1-Flash raising AIME/HMMT 2026 from 64% to 71% with a self-written code layer
The lab calls this mechanistic intervention: an agent reads internal state and changes data, objectives, update rules or architecture while …
Agent MechSuits: white-box safety steering for multi-turn CLI agents by intervening in a 10-dimensional subspace of one layer
A linear probe reads step-level hidden states and, when it flags risk, steers a 10-dim subspace without changing weights. On the MAS benchma…
HeliosLM: open Traditional Chinese (Taiwan) chat, speech recognition and safety models released on Hugging Face
Chat and safety models are based on Gemma 4 E4B-it; ASR models on Whisper large-v3. Self-reported CER on Common Voice zh-TW falls from 15.25…
知乎热榜：54 岁大哥自学 AI 短剧登顶红果榜，普通人真能靠 AI 翻身吗？
这条问题以约 980 万热度出现在知乎热榜，讨论的是有人用 AI 做短剧登上红果榜、团队成员一年还清债务的传闻，以及普通人有没有可借鉴的路径。它更多是社会讨论而不是一手技术报告，但对想观察国内 AI 短剧商业化情绪的人，可以看出讨论已经从「能不能做」转到「普通人能不能靠它翻身」。
知乎热榜：张艺谋称 AI 带来「平视感」，「普通人都是导演、编剧」，怎么看？
热度约 99 万。讨论围绕张艺谋关于 AI 技术飞跃降低电影门槛、带来平视感的表态。这是文化评论而不是产品发布，但和近期国内 AI 短剧、AI 影像工具进入大众创作的讨论连在一起，可以看出影视圈对「人人都能当导演」这句话的分化反应。
Range 的理财规划师 Matt Ragone：没有编程背景的规划团队，怎样自己做出每天在用的 Agent 和技能
在 AI Engineer 的演讲里，财富管理机构 Range 的规划师讲他们的「选、教、打包、测、上线」五步：先挑每天重复的工作（现金流分析、保险评估），用真实案例和公司标准化文档教 Agent，打包成技能时让它先提问、并禁止直接发给客户，上线前用 5 到 10 个真实案例（含…
Tanmay Sah 和 Dolly Sah：编码代理做对任务，不代表它能撤回自己的错误
他们提出 Undo Bench：在 2880 次成对实验里，代理正常成功率约 84%，但一旦引入故障，这些原本能成功的实验里只有约 47% 能恢复；对「请求超时但外部操作其实已完成」这类情况，盲目重试在约 54% 的实验里造成了重复副作用。Git revert 只能撤回仓库里的改…
Addy Osmani: Your coding agents made you a manager
The Claude Code engineer argues that spending the day with coding agents is already management: choosing tasks, writing briefs, delegating, …
Rob Miles: I never want to use your AI assistant on your website. I want to use my AI assistant on your website
The AI safety researcher’s short post argues companies should let people bring their own agents rather than forcing a built-in assistant, a …
OpenAI’s Tibo: comparing GPT 6.1 Terra with Opus 5.5 is “pretty wild”; it isn’t meant to compete in the big leagues
Replying to a preference poll, Codex lead Thibault Sottiaux said Terra is a great model but not positioned against Opus 5.5, and that OpenAI…
The inventor of the bitter lesson has a bitter lesson for LLM scaling enthusiasts
The Decoder reports Richard Sutton’s latest reminder that raw compute is not enough, adds Jerry Tworek’s critique of a generation “reward-ha…