# Ziyu 的 AI 影片工作室筆記

> AI 影片工作室與軟體工程的實作筆記——全本地生成、代理協作、雙平台發佈

Markdown blog covering AI tooling, video-generation engineering, and software practice.

## Posts

- [我的 AI 影片頻道總覽:四個系列,全部在一台 MacBook 上生成](https://ziyu4huang.github.io/blogs/posts/2026-10-04-ai-video-channel-guide/): 本站對應的影片頻道總索引——《跟著致敬片學 AI 分鏡》教學系列、三個致敬/復刻系列,每一集的 B站 與 YouTube 連結都在這,持續更新。
- [h3c 客製全紀錄:從 antirez 的 h3.c 到量產 Metal 影片引擎](https://ziyu4huang.github.io/blogs/posts/2026-10-04-h3c-engine-custom-build/): C11 + Metal 直讀 safetensors、三顆引擎樹的版本紀律、七項 fork 客訂——控制 tokens、音訊 sigma schedule、pruned AdaLN、runtime LoRA 檢查——加上互動佇列,我們把 MiniMax H3 鍛成一顆在 MacBook 上量產有聲影片的本機引擎。
- [生產宣言:純本地、零雲端的指揮式 AI 影片工作室](https://ziyu4huang.github.io/blogs/posts/2026-10-04-local-video-studio-philosophy/): 在一台 MacBook Pro M5 Max 上,人只下命令,AI 自主完成規劃、prompt、渲染、QC、修復到發佈——這是我經營本地 AI 影片工作室的三條鐵則與配套原則。
- [MiniMax H3 品質基準測試:在 Apple Silicon 上找到影片生成的甜點](https://ziyu4huang.github.io/blogs/posts/2026-10-04-minimax-h3-quality-benchmark/): 從 20 步到 30 步、turbo LoRA 到完整 66GB transformer、608×352 到 720p——一份在 M5 Max 上實測的 MiniMax H3 影片品質階梯,附七個 Phase C 實驗的判決紀錄。
- [Fubon Neo API: A Developer's Guide to Taiwan Stock Market Data](https://ziyu4huang.github.io/blogs/posts/2026-04-09-fubon-neo-api-taiwan-stock/): A developer tour of the Fubon Neo API for Taiwan market data: auth, quotes, rate limits, and TypeScript patterns for turning raw ticks into usable signals.
- [Building a Taiwan Stock Market Explainer Video with Remotion + Claude Code](https://ziyu4huang.github.io/blogs/posts/2026-04-08-remotion-claude-code-taiwan-stock-video/): Building a Taiwan stock explainer video end-to-end with Remotion and Claude Code: data pipeline, scene composition, narration, and the render tricks that made it watchable.
- [Nano Banana: Google's AI Revolution in Image Generation and Editing](https://ziyu4huang.github.io/blogs/posts/2025-10-06-nano-banana-ai-revolution/): Hands-on with Google Nano Banana: what it generates well, how its edit mode compares to plain re-prompting, and a gallery of generation and edit examples.
- [Why We Need Claude Code SDK: Pros, Cons, and the Future of AI Development Tools](https://ziyu4huang.github.io/blogs/posts/2025-10-06-why-we-need-claude-code-sdk/): Where the Claude Code SDK shines, where it fights you, and how to choose between SDK, CLI and raw API when automating agent workflows.
- [Configure Z.ai with Claude Code: Complete Guide to GLM-4.6 Integration](https://ziyu4huang.github.io/blogs/posts/2025-10-04-configure-zai-claude-code/): Step-by-step guide to wiring Z.ai GLM-4.6 into Claude Code: base URL, auth token, model mapping, and the pitfalls (context limits, MCP output caps) hit along the way.
- [Introducing BMAD-METHOD: A Universal AI Agent Framework](https://ziyu4huang.github.io/blogs/posts/2025-10-04-introducing-bmad-method/): What the BMAD-METHOD multi-agent framework actually does, how its analyst-architect-dev-QA role chain maps onto real builds, and when it beats a single well-prompted session.