The second part of the context engineering series. Part 1 covered the foundations: what an LLM is, how it becomes an agent, and the seven layers of configuration (CLAUDE.md, hooks, MCP, skills, plugins) that make it production-ready. This part picks up where we left off — with subagents, agent teams, worktree isolation, and the context management machinery that keeps it all running as sessions scale. We close with a look at how the broader open-source ecosystem is converging on the same orchestration patterns from different directions.

A practical, end-to-end guide to building a custom MCP server that turns your documentation site into structured tools for LLM agents. Covers the llms.txt standard, FastMCP, caching, search, packaging, and LLM agents integration. Written for teams that have internal docs (Confluence pages, Markdown files) and want their agents to read them instead of hallucinating.

A practical guide to LLM agents, from foundational concepts to building a production-ready agent configuration. Part 1 of a two-part series: what an LLM actually is, how it becomes an agent, and the layered system (CLAUDE.md, hooks, MCP, skills, plugins) that makes it useful in practice. Part 2 covers subagents, agent teams, and more advanced topics. Written for everyone, regardless of technical background.

和 Gemini 聊爽了，以后直接开这么个栏目得了，类似 @AI-Talk 之前做的「艾尔文团长对话谏山创」，不过性质上可能更接近《理想国》。毕竟时间有限，就像我现在 > 95% 的代码都是 LLM 生成的一样，AI 时代的写作思路也应该转变了。好比中泽工「I/O」的文本填充也不是自己亲自完成的，以后我就是我文章的「总策划」，而不必要是「写手」。

故事的开始是今天下班后随手打了一段文字，准备发个动态，分享下我此刻尽管百般劳累、但在看到工资条后阴霾仿佛一扫而空的喜悦心情。正所谓「打工人一天最爽的时候是下班时，一周最爽的时候是周末开始时，一个月最爽的时候是发工资时，一年最爽的时候是发年终奖时」——今天可谓是连占三样。然而写到一半时我自我审视了下，即使作为无需费心遣词造句的日常吐槽，这种字里行间不经意流露的高傲态度也很让人不适。这最多只能作为私人的日记，而不适合在公开场合发表。本质上除了满足自己表达欲的宣泄，对他人而言这种文字有什么看的意义呢？于是我决定拉倒，反正文字在写完的瞬间就已经达到了原来的目的。

但我突然冒出一个想法：如果我把这段文字交给 LLM，让他来对我进行评判、精神分析，会得到怎样的结果？于是就诞生了这篇完全非我本人写作、以对话形式呈现的「文章」。结果有些超出预期（否则我也不至于特意整理成文了）。现在的 LLM 水平真挺了得，我也推荐大家拿自己平时的发言试一试，将自己视作文本做一做「症候阅读」。平日里难以觉察的意识形态正隐匿在文本的间隙中。