You can now use Claude as an orchestration agent for long-form video generation.
From one text prompt, it can direct various tools and combine outputs - like Nano Banana Pro for images, Hailuo 2 for video, and ffmpeg to stitch.
Watch it construct my bookstore x coffee shop 👇
Today, Veo 3.1 Ingredients to Video is getting more expressive + portrait mode has arrived (we heard you!)
You can now create vertical clips with livelier movement and better control directly in the @Geminiapp, @YouTube & Google Vids.
All this plus state-of-the-art upscaling to 1080p and 4K also coming to @flowbygoogle, the Gemini API and Vertex AI.
Today we are shipping expanded URL support 🔗 in the Gemini API!! You can now pass signed or public URLs directly to Gemini and it will fetch images or PDFs, along with support for a new integration with Google Cloud Storage allowing you to keep you data in one place!!
我写了一个自动发微信公众号的 Skill，操作 Chrome，支持图文和文章，会记住登录状态，不需要每次登录。
文章的话，提供markdown文档本地地址，会自动帮你把 Markdown 转 HTML（可选风格较少），配图会一张张粘贴进编辑器，不需要手动上传，封面图、原创设置暂时不支持，建议把封面图放文章内容。
技术上我没有使用 PlayWright MCP，因为这玩意儿太费 Tokens 了，而是用的 Chrome CDP (Chrome DevTools Protocol) 是一个允许外部程序通过 WebSocket 与 Chrome（以及 Edge、Opera 等基于 Chromium 浏览器的内核）进行通信的底层调试协议。
都是脚本操作，不怎么费 Token。
图文的话需要告诉图片地址、标题和内容，可以自动上传图片，填写标题和内容。
所有操作都不会发布，只是帮你生成草稿。
需要 Claude Code 或者其他支持 Skills 的 Agent，需要 Nodejs 运行环境（但如果你装了 Claude Code 应该就支持 Node）
Skill 地址：github.com/JimLiu/baoyu-s…
安装说明：
github.com/JimLiu/baoyu-s…
这是我分享的 Skill 之一，还有一些其他 Skills，注意其中 gemini-web 的 skill 可以帮你用你的 Gemini 账号画图，需要自己登录一下。不保证它的稳定性和安全性，不过我自己也在用。
推荐王老师的教程，另外刚才测试了一下王老师写的自动发布 X 文章的 Skill，真的是强大，而且给我很大启发，理论上来说基于这个思路可以做一个发布微信公众号或者其他平台的 Skill。
原理是用脚本控制浏览器，用剪贴板把文字和图片粘贴到编辑器，最有创意的是，根据文字定位到图片要插入的位置👍
Things to keep in my when building Agents:
- Use Unix basics, file system and a command line interface (Bash).
- All tools and sub-agents should share the same file system to read and write each other's work.
- Use code generation for non-coding tasks, like analyzing data or chaining tool calls.
- Verification is critical, Make the agent run a script or check for errors before finishing a task.
- Use sub-agents to handle specific tasks and return only the final answer to avoid confusing the main agent.
- Avoid loading huge files into the context. Write scripts/commands to search and extract only the specific data needed.
- Reading the agent's traces to understand exactly why and where it made a mistake.
- Avoid semantic search for precise data. A query for structured data is often more reliable.
- Design your agent to handle errors by reading error messages.
Is Agentic RAG worth it?
RAG systems have evolved from simple retriever-generator pipelines to sophisticated workflows. It remains unclear when to use Enhanced RAG (fixed pipelines with dedicated modules) versus Agentic RAG (LLM orchestrates the entire process dynamically).
This research provides the first empirical comparison.
Enhanced RAG adds pre-defined components to address specific weaknesses: routers to determine if retrieval is needed, query rewriters to improve alignment, and rerankers to refine document selection. The workflow is fixed and manually engineered.
Agentic RAG takes a different approach. The LLM decides which actions to perform, when to perform them, and whether to iterate. No extra components beyond the basic knowledge base, retriever, and generator. The model controls everything.
The researchers evaluated both paradigms across four dimensions on QA and information retrieval tasks.
User intent handling: Agentic slightly outperforms Enhanced on most tasks, but Enhanced wins decisively on FEVER (+28.8 F1 points), where the agent often retrieves unnecessarily.
Query rewriting: Agentic RAG achieves 55.6 average NDCG@10 compared to 52.8 for Enhanced, showing the agent can adaptively rewrite queries when beneficial.
Document refinement: Enhanced RAG with reranking (49.5 NDCG@10) outperforms Agentic (43.9). Dedicated reranker modules beat iterative retrieval attempts.
Agentic RAG is far more sensitive to model capability. With weaker models, Enhanced RAG maintains stability while Agentic performance degrades significantly.
Cost analysis reveals Agentic RAG requires 2-10x more computation time and tokens due to multi-step reasoning.
The choice between Enhanced and Agentic RAG depends on your constraints. Enhanced RAG offers predictability, lower costs, and stability with weaker models. Agentic RAG provides flexibility but requires stronger models and more compute.
Paper: arxiv.org/abs/2601.07711
Learn to build effective Agentic RAG systems in our academy: dair-ai.thinkific.com/pages/courses
Vibe Coding 拼到后面又回到了传统软件工程和 coding。
同样是写提示词，专业的人一步到位，清晰描述，需求是什么，输入输出是什么，可以参考什么代码，可能需要什么文档，去哪里找，写完了怎么验证。
不懂的人怎么也描述不到重点：“我想要功能强大的 Coding Agent”，别说 AI 不懂，连他们自己都不懂自己要什么。
不懂代码的话，刚开始的时候，代码少功能简单，生成后就能运行，代码多了你就只知道跑起来很卡，各种 Bug，但是都不知道怎么跟 AI 描述你遇到的问题。
本地跑跑其实还好，部署到生产环境给别人用更可怕，API Key 写在客户端还以为别人不知道。
懂代码的话，每次 AI 生成的代码，不是无脑接受，而是审查代码，看是不是满足需求、有没有安全性能问题、是不是好维护。出问题知道怎么重现，怎么描述问题，怎么调试。
如果你只是做原型，或者自己用的小脚本其实还好，如果是做对外的产品，还是要慎重，有敬畏之心。
试了一下claude的cowork
只得感叹别人vibe coding的水平和我现在的水平差距实在太大了
我还处于按起葫芦浮起瓢的修bug状态
人家已经能做一套完整丝滑的产品体验了
应该去claude偷师一下😅
More Gemma! Meet TranslateGemma, a new collection of open translation models built on Gemma 3 designed for high-performance communication.
- Available in 4B, 12B, and 27B parameter sizes.
- Evaluated on 55 languages using the WMT24++ dataset.
- 12B model outperforms the Gemma 3 27B baseline.
- 4B model optimized specifically for mobile and edge deployment.
huggingface.co/collections/go…
GitHub Copilot now remembers what it learns across your workflow.
➡️ Copilot coding agent, CLI, and code review share knowledge automatically
➡️ Memories include citations that get verified in real-time before use
➡️ Outdated information self-heals as agents discover contradictions
➡️ 7% higher PR merge rates in testing, 2% better code review feedback
Available now in public preview for all paid Copilot plans. Deep dive below ⬇️
github.blog/ai-and-ml/gith…
This week, we released a pair of Gemma updates that developers can use to innovate across the healthcare and translation spaces:
— MedGemma 1.5 4B: Our updated open medical AI model improves performance in 3D imaging and reasoning over medical records. We also kicked off the MedGemma Impact Challenge on @Kaggle to inspire development of AI applications that can help patients
— TranslateGemma: This collection of multimodal open translation models is built on Gemma 3 and enables developers to build translation tools across 55 language pairs. The models are efficiently compact (available in 4B, 12B, and 27B parameter sizes), achieving high-fidelity translation quality with lower latency
You can access both Gemma models on @huggingface, MedGemma on Vertex AI, and TranslateGemma on Kaggle. To learn more about Gemma and our latest breakthroughs from @GoogleDeepMind, visit deepmind.google!
