Blog 博客

Documentation is write-once-read-maybe-update-never. A Skill is a persisted human-agent contract — agents write them, humans curate them, and knowledge stays alive in the loop. Why Skill marketplaces will inevitably fail, and why everyone needs their own Skill library.

文档是 write-once-read-maybe-update-never。Skill 是 human-agent contract 的持久化形式——agent 写、人 curate、知识在循环中活起来。Skill marketplace 为什么必然失败，每个人为什么需要自己的 Skill library。

Prompt engineering hits a cliff in long-run tasks. Not because your prompts are bad — because attention decay, context rot, and drift accumulation are physics. Harness engineering is the complete answer.

Prompt engineering 在 long-run task 面前断崖式失效。不是因为 prompt 写得不好，是因为 attention decay、context rot、drift accumulation 三重物理极限。Harness engineering 是完整的答案。

From the Cognition Ladder to the Capability × Validation Matrix: New thoughts on the Agent era, one year later

从认知阶梯到能力×验证矩阵：一年之后，我对 Agent 时代的新思考

LLMs satisfice by default. Harness engineering constrains agent behavior with structure, not prompts. A four-layer architecture and how we landed it in OPC.

LLM 的默认行为是 satisfice，不是 optimize。Harness engineering 用结构约束 agent 行为，而不是用 prompt 期望 agent 行为。一个四层架构的设计理念和 OPC 落地。

I built a Claude Code skill that dispatches 11 specialist AI agents to review code from different perspectives. Here's an honest comparison with just asking Claude directly.

我做了一个 Claude Code skill，用 11 个 AI 专家并行 review 代码。这是一个诚实的对比：OPC vs 直接问 Claude。

AI agents start from scratch every session. I built memex — a Zettelkasten-based persistent memory system that needs no vector DB. Here's the design philosophy and trade-offs behind it.

AI agent 每次新 session 都从零开始。我用 Zettelkasten 方法构建了 memex —— 一个不需要 vector DB 的 agent 持久记忆系统，背后的设计哲学和技术取舍。

Superpowers 是一套 Claude Code 的 skill 插件。深入分析其设计哲学：从流程强制到 Zero Trust 架构，再到用 AI 弱点约束 AI 的元层机制。

Superpowers is a set of skill plugins for Claude Code. A deep dive into its design philosophy — from enforced workflows to Zero Trust architecture, to a meta-layer mechanism that uses AI's weaknesses to constrain AI.