飞书办了个 Skill 比赛，有个赛道叫邪修。我把 29 天造的算命产品里三个东方占卜模块抠出来，打包成一个赛博半仙 skill。这一篇是过程记录，也是 AX Skill Workshop 系列开篇。

Took the three Eastern divination engines from PanPanMao and packaged them into a single agent-portable skill — algorithm engine + knowledge base + persona prompt, three folders deep. The pattern transfers to any product with real domain knowledge.

先是裸跑 Claude Code。再是 GSD 多 agent 仪式。再是 superpowers 的 skill 库。最后在上面包了一个自己的 orchestrator。每一版崩盘都来自一个具体 session。最后我才想明白：solo dev 的 harness 跟团队的不是一个物种，而且真正的答案不是选一个默认引擎——是按工作形状路由。

Vanilla Claude Code. Then GSD's multi-agent ceremony. Then superpowers' skill system. Then a custom orchestrator on top. Each rewrite came from a specific session that broke. What I learned: solo dev harnesses and team harnesses are different species, and the right answer isn't a default engine — it's routing by work shape.

AX 的 AI 雷达每天从 50 多个信源抽约 500 条内容，挑 15 条给我。选题不靠 prompt，靠一份能版本化、能 diff 的 Markdown 编辑策略——一个 Agent 从反馈里迭代它。讲这把雷达的架构取舍和关键设计决策。

AX Radar filters ~500 stories a day from 50+ sources down to about 15. Editorial judgment isn't a prompt — it's a versionable, diff-able Markdown file that a Claude Agent iterates against user feedback. A breakdown of the architecture and the design calls.

上一篇讲的是我用 AX 雷达的一手体验。这一篇讲我给它加了 HTTP API、MCP server 和一份 Claude Skill——让 Claude 自己能查、能搜、能保存。两张脸共用一个后端；sha256 不是 bcrypt；pgvector 早就在了；Skill 才是放 domain 知识的地方。

Part 1 covered AX Radar from my side as the operator. Part 2 covers what I shipped this week: an HTTP API, an MCP server, and a Claude Code skill that let Claude read, search, and mutate the radar directly. Two surfaces, one backend. Why sha256 beats bcrypt here. Why semantic search rode the existing pgvector index. Why the skill is where domain knowledge lives.

AI 随想第十篇说代码学会了自我进化。那是观察。这篇是动手。

In Part 10 of Agentic AI Thoughts, I wrote that code had learned to evolve itself. That was the observation. This is the implementation.

三条需求曲线、一条成本下降线，以及为什么算力是新石油

Three demand curves, one deflation trend, and why compute is the new oil

Mio 有 25 个角色，每个都有深度人格——但所有交互都是一对一聊天。这是一种浪费。如果可可、了空、陆霆坐在一起玩狼人杀呢？如果角色创建不是填表，而是聊天呢？

Mio has 25 personas with deep, distinct personalities — but every interaction is one-on-one chat. That's a waste. What happens when Coco, Liao Kong, and Lu Ting sit down at a Werewolf table? And what if creating a character wasn't a form, but a conversation?

每个 AI 伴侣产品都会撞上同一面墙：聊着聊着就没劲了。这不是提示词写得不够好，不是模型不够聪明——是纯聊天这种形态，从结构上就缺了两个维度。我在做 Mio 和 Lumi 的过程中，慢慢看清了这个问题。

Every AI companion product hits the same wall: pure chat gets boring. Not because the models aren't smart enough or the prompts aren't good enough — because chat as a format is structurally missing two dimensions that real relationships need. Here's what building Mio and Lumi taught me about this ceiling.

棉花娃娃、Labubu、AI 伴侣、乌托的 OC——一整代人正在把养育本能投射到虚拟生命上。这不是亚文化，是一个百亿级的情感迁移。而真正的问题是：如果人一定会跟虚拟生命建立深度关系，那这个生命该是什么？

Cotton dolls, Labubu, AI companions, virtual OCs — an entire generation is redirecting parental instinct toward virtual beings. This isn't a subculture. It's a multi-billion dollar emotional migration. The real question: if people will inevitably form deep attachments to virtual beings, what should those beings be?

乌托把荣格八维拆成 8 个 agent，用权重阈值触发 MBTI 切换。Mio 和 Lumi 让性格从对话中涌现——你聊什么，它就长成什么。两条路，各有各的好。但有一个瞬间让我改变了看法。

One team splits Jung's 8 cognitive functions into 8 agents and triggers MBTI shifts with weight thresholds. We let personality emerge from conversation — you talk, it becomes. Two philosophically different approaches to giving AI a soul. And one unexpected moment that changed how I think about both.