Hi, I'm Jingxia 👋🏻 你好，我是璟夏 👋🏻

• Documentation Rots. Skills Stay Alive.

  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.

  --
• 文档写完就落灰了，Skill 越用越活

  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。

  --
• Harness Engineering II: When Tasks Go From Minutes to Hours

  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.

  --
• Harness Engineering II: 当任务从分钟级变成小时级

  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。

  --
• The Continued Evolution of the Agent Era

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

  --
• Agent 时代的持续进化

  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

  --
• Harness Engineering: Don't Hope Your AI Behaves — Build a Structure Where It Can't Not

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

  --
• Harness Engineering: 别期望 AI 行为正确，构造一个结构让它想不正确都难

  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。

  --
• OPC — Your AI Review Team in One Slash Command

  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.

  --
• OPC — 一个斜杠命令召唤你的 AI Review 团队

  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。

  --

• Loading... 加载中...

• 3/10/2026 #AI

  Countering AI laziness patterns: treating lower bounds as targets (>=8 pages → writes exactly 8); skipping process to jump to conclusions; template filling instead of thinking; choosing the narrowest interpretation of vague requirements to claim completion; answering from memory when deep search is needed.

  Fighting AI laziness is really fighting human laziness: when you rush to get things done sloppily, so does your AI.

  3/10/2026 #AI

  Countering AI laziness patterns: treating lower bounds as targets (>=8 pages → writes exactly 8); skipping process to jump to conclusions; template filling instead of thinking; choosing the narrowest interpretation of vague requirements to claim completion; answering from memory when deep search is needed.

  Fighting AI laziness is really fighting human laziness: when you rush to get things done sloppily, so does your AI.

• 2026/3/10 #AI

  对抗AI的惰性模式：数量约束取下限（>=8页只写8页）；跳过过程给结论；模板填充代替思考；越模糊的需求越倾向做最少的事claim完成；需要深度搜索的问题凭记忆回答。

  对抗AI的惰性，其实就是对抗人的惰性：你着急把事情很粗糙的做完，你的AI也是。

  2026/3/10 #AI

  对抗AI的惰性模式：数量约束取下限（>=8页只写8页）；跳过过程给结论；模板填充代替思考；越模糊的需求越倾向做最少的事claim完成；需要深度搜索的问题凭记忆回答。

  对抗AI的惰性，其实就是对抗人的惰性：你着急把事情很粗糙的做完，你的AI也是。

• 3/10/2026 #thinking

  Every time a new product or open-source project drops, you see a crowd beating their chests in regret, and an even bigger crowd applauding. The cognition gap is exactly this — those one level below see nothing but great ideas; those who’ve reached that level see “idea is same and cheap,” and if you move too slowly, you miss every hot opportunity.

  3/10/2026 #thinking

  Every time a new product or open-source project drops, you see a crowd beating their chests in regret, and an even bigger crowd applauding. The cognition gap is exactly this — those one level below see nothing but great ideas; those who’ve reached that level see “idea is same and cheap,” and if you move too slowly, you miss every hot opportunity.

• 2026/3/10 #thinking

  每次一个新产品或者开源项目出来，能看到一堆人捶胸顿足，然后更多的人拍手叫好。认知的鸿沟莫过于此——认知水平不够下一级的感受到的都是好主意，认知水平到了的感受到的都是 idea is same and cheap，动作慢了啥都吃不上热的。

  2026/3/10 #thinking

  每次一个新产品或者开源项目出来，能看到一堆人捶胸顿足，然后更多的人拍手叫好。认知的鸿沟莫过于此——认知水平不够下一级的感受到的都是好主意，认知水平到了的感受到的都是 idea is same and cheap，动作慢了啥都吃不上热的。

• 3/10/2026 #AI

  Intuition is the highest form of rationality. AI will render 90% of our skills useless — but it can amplify the remaining 10% by 10x, 100x, 1000x. Intuition is surely one of the most important traits in that 10%.

  3/10/2026 #AI

  Intuition is the highest form of rationality. AI will render 90% of our skills useless — but it can amplify the remaining 10% by 10x, 100x, 1000x. Intuition is surely one of the most important traits in that 10%.

• 2026/3/10 #AI

  直觉，是最高级的理性。AI 会让我们90%的技能毫无用处，但是，可以10倍100倍1000倍放大我们剩下的那10%。直觉，一定是那10%里面非常重要的一个特征。

  2026/3/10 #AI

  直觉，是最高级的理性。AI 会让我们90%的技能毫无用处，但是，可以10倍100倍1000倍放大我们剩下的那10%。直觉，一定是那10%里面非常重要的一个特征。

• 3/10/2026 #agent

  #OpenAI Operator

  • Spilled the agent dividend from programmers to information workers, but hard to break out further
  • Delivered the “one carbon, many silicon” scenario beautifully
  • Efficiently improved agent anthropomorphism in a simple way
  • Showed us what’s possible once you have most of a person’s permissions and information — the ceiling is now clear to many, more thorough than Anthropic’s incremental approach (grateful for open source — able to go viral without user backlash)

  Opportunities

  • How to spill further to a broader audience: a lightweight on-device app + a set of intuitive scenarios + a cloud host
  • Monetization is a major open question
  • Best practices for building OPC (One Person Company) on Operator
  • Keep making shovels — used by decentralized/fragmented scenario-specific applications
  • Many carbon, many silicon: how to go from OPC to Multi Person Company
  3/10/2026 #agent

  #OpenAI Operator

  • Spilled the agent dividend from programmers to information workers, but hard to break out further
  • Delivered the “one carbon, many silicon” scenario beautifully
  • Efficiently improved agent anthropomorphism in a simple way
  • Showed us what’s possible once you have most of a person’s permissions and information — the ceiling is now clear to many, more thorough than Anthropic’s incremental approach (grateful for open source — able to go viral without user backlash)

  Opportunities

  • How to spill further to a broader audience: a lightweight on-device app + a set of intuitive scenarios + a cloud host
  • Monetization is a major open question
  • Best practices for building OPC (One Person Company) on Operator
  • Keep making shovels — used by decentralized/fragmented scenario-specific applications
  • Many carbon, many silicon: how to go from OPC to Multi Person Company
• 2026/3/10 #agent

  #OpenAI Operator

  • 把原来程序员基于 agent 的红利溢出到 informational worker 了，但是很难继续出圈
  • 把单 carbon 多 silicon 这个场景水灵灵的实现了
  • 用简单的方式很高效率地提升了 agent 的拟人化能力
  • 告诉我们拿到人的绝大部分权限和信息之后，能够做到啥，天花板让很多人一眼看清了，比 Anthropic 渐进式的来得更彻底（感恩开源，能够在不被用户反噬的情况下迅速出圈）

  机会点

  • 怎么继续溢出给更广泛的群体：一个轻便的端侧 app + 一组符合直觉的场景 + 一个云端主机
  • 如何收费是大问题
  • 基于 Operator 打造 OPC 的各类最佳实践
  • 继续做铲子，被去中心化/碎片化的各种场景化应用使用
  • 多 carbon 多 silicon：如何从 OPC 走向 Multi Person Company
  2026/3/10 #agent

  #OpenAI Operator

  • 把原来程序员基于 agent 的红利溢出到 informational worker 了，但是很难继续出圈
  • 把单 carbon 多 silicon 这个场景水灵灵的实现了
  • 用简单的方式很高效率地提升了 agent 的拟人化能力
  • 告诉我们拿到人的绝大部分权限和信息之后，能够做到啥，天花板让很多人一眼看清了，比 Anthropic 渐进式的来得更彻底（感恩开源，能够在不被用户反噬的情况下迅速出圈）

  机会点

  • 怎么继续溢出给更广泛的群体：一个轻便的端侧 app + 一组符合直觉的场景 + 一个云端主机
  • 如何收费是大问题
  • 基于 Operator 打造 OPC 的各类最佳实践
  • 继续做铲子，被去中心化/碎片化的各种场景化应用使用
  • 多 carbon 多 silicon：如何从 OPC 走向 Multi Person Company
• 3/10/2026 #AI

  Superpowers doesn’t persuade AI at a logical level — it manipulates AI’s token generation distribution at a statistical level. Using AI’s weaknesses to constrain AI.

  3/10/2026 #AI

  Superpowers doesn’t persuade AI at a logical level — it manipulates AI’s token generation distribution at a statistical level. Using AI’s weaknesses to constrain AI.

• 2026/3/10 #AI

  Superpowers 不是在逻辑层面说服 AI，而是在统计层面操控 AI 的 token generation distribution。用 AI 的弱点约束 AI。

  2026/3/10 #AI

  Superpowers 不是在逻辑层面说服 AI，而是在统计层面操控 AI 的 token generation distribution。用 AI 的弱点约束 AI。

• OPC

  One Person Company — a digraph-based task pipeline with independent multi-role evaluation for AI agents.
• Memex

  Zettelkasten-based persistent memory for AI coding agents. No vector DB — just markdown + git.
• SuperCrew

  Superpowers skills, hooks, commands, and agents for Claude Code.

📬 订阅更新 📬 Subscribe

不定期更新，聊 agent 架构、harness engineering 和 AI 工程实战。不灌鸡汤，只写踩过的坑。 Occasional updates on agent architecture, harness engineering, and real AI engineering lessons. No hype, just hard-won lessons.

由 Buttondown 提供 · 随时退订 Powered by Buttondown · Unsubscribe anytime

Let's Connect 联系我

• github /
• twitter /
• linkedin /
• 微信 /
• xingjingx@gmail.com