关于 AI、Agent、产品与判断的持续笔记。https://roarkist.com/zh-cnhttps://roarkist.com/series/agent-foundations/08-build-research-agent/https://roarkist.com/series/agent-foundations/08-build-research-agent/用检索、阅读、记录与复核组成一条可观察的最小研究流程。Thu, 10 Sep 2026 00:00:00 GMThttps://roarkist.com/series/agent-foundations/07-observe-and-evaluate/https://roarkist.com/series/agent-foundations/07-observe-and-evaluate/不要只看最终答案，还要检查路径、工具调用、成本与失败恢复。Wed, 09 Sep 2026 00:00:00 GMThttps://roarkist.com/series/agent-foundations/06-human-checkpoints/https://roarkist.com/series/agent-foundations/06-human-checkpoints/把人工确认放在高影响、难撤回和信息不足的决策之前。Tue, 08 Sep 2026 00:00:00 GMThttps://roarkist.com/series/agent-foundations/05-planning-before-action/https://roarkist.com/series/agent-foundations/05-planning-before-action/规划适合降低长任务的不确定性，但不应成为每次行动前的固定仪式。Mon, 07 Sep 2026 00:00:00 GMThttps://roarkist.com/series/agent-foundations/04-state-memory-context/https://roarkist.com/series/agent-foundations/04-state-memory-context/分清三者，才能避免让聊天记录承担所有信息管理工作。Sun, 06 Sep 2026 00:00:00 GMThttps://roarkist.com/series/agent-foundations/03-tools-as-contracts/https://roarkist.com/series/agent-foundations/03-tools-as-contracts/工具定义得越清楚，Agent 的选择、执行和错误恢复就越可靠。Sat, 05 Sep 2026 00:00:00 GMThttps://roarkist.com/series/agent-foundations/02-agent-loop/https://roarkist.com/series/agent-foundations/02-agent-loop/用一个可观察、可停止的循环理解 Agent 如何推进任务。Fri, 04 Sep 2026 00:00:00 GMThttps://roarkist.com/series/agent-foundations/01-agent-vs-chatbot/https://roarkist.com/series/agent-foundations/01-agent-vs-chatbot/关键不在拟人化，而在它能否围绕目标持续采取行动并检查结果。Thu, 03 Sep 2026 00:00:00 GMThttps://roarkist.com/blog/every-product-needs-an-editor/https://roarkist.com/blog/every-product-needs-an-editor/产品不只需要增加能力，也需要有人持续决定什么应该留下。Wed, 02 Sep 2026 00:00:00 GMThttps://roarkist.com/blog/best-ai-workflow/https://roarkist.com/blog/best-ai-workflow/Agent 能提高局部速度，但真正决定结果的是上下文、检查点与责任边界。Mon, 24 Aug 2026 00:00:00 GMThttps://roarkist.com/blog/fast-implementation-slow-decisions/https://roarkist.com/blog/fast-implementation-slow-decisions/AI 缩短了实现时间，但没有替我们承担方向错误的成本。Sun, 09 Aug 2026 00:00:00 GMThttps://roarkist.com/blog/software-room-for-judgment/https://roarkist.com/blog/software-room-for-judgment/自动化应当减少机械负担，而不是把复杂决定伪装成一个确定答案。Sun, 26 Jul 2026 00:00:00 GMThttps://roarkist.com/blog/three-skills-relearning/https://roarkist.com/blog/three-skills-relearning/澄清问题、缩小范围、完成复盘，是我在每个项目里都会重新学习的事。Wed, 08 Jul 2026 00:00:00 GMThttps://roarkist.com/blog/fuzzy-idea-to-product/https://roarkist.com/blog/fuzzy-idea-to-product/从一句含糊抱怨开始，经过原型、删减和真实使用，才逐渐看见产品。Thu, 18 Dec 2025 00:00:00 GMThttps://roarkist.com/blog/learning-vibe-coding/https://roarkist.com/blog/learning-vibe-coding/不会因为 AI 写代码就跳过工程，而是用更短的反馈循环学习工程。Mon, 06 Oct 2025 00:00:00 GMThttps://roarkist.com/blog/organizing-repeated-work/https://roarkist.com/blog/organizing-repeated-work/当同一种解释和检查不断出现时，它可能已经值得成为一套可复用流程。Sun, 15 Jun 2025 00:00:00 GMT