#编码

Day 6/ Reset

ICYMI: Composer predictions are now live for all Pro users in Codex. It allows codex to analyse your entire conversation and project context to suggest the next best prompt. In the past 2 weeks, it’s successfully predicted 36% of my next steps & about 11% more with minor changes to the suggestion. Give it a shot ;)

🗞️ if your team is too busy doing their 'normal job' to experiment with AI, you're preparing them to be replaced https://ghuntley.com/replaced/ In this post, I guess this's the last time I'll say it. AI use is no longer optional if you wish to be employed. What follows is advice for current employees and hiring managers on how to hire in 2026 to source AI-first candidates.

Now in beta: composer predictions in Codex for Pro users. Codex can now suggest your next message based on your conversation and how you talk to it. One of the most loved new features we've ever tested internally.

Now in beta: composer predictions in Codex for Pro users. Codex can now suggest your next message based on your conversation and how you talk to it. One of the most loved new features we've ever tested internally.

@StepFun_ai just announced Step 5 Preview, and it's FREE in Kilo for a week: • 600B parameters, 27B active per token • 1M-token context + vision • highlights: strong coding and finance performance at lower cost. Try it and let us know what you think!

I am now managing 90% of my coding work through Grok @bot. And it's amazing.

Should you /grill-with-docs, /wayfinder, or just one shot the code? 1. If the diff is tiny, DON'T grill /grill-with-docs saves you from situations where the agent built the wrong thing. But if the thing it's building is tiny, that cost drops to nothing. So one-shotting is absolutely fine when you're only making small changes. 2. Never start with /wayfinder A failure mode of /wayfinder is that you start grilling and the solution turns out to be way simpler than you thought. Why is this bad? Because you've created the map and tickets, and you realise you don't need them. So if the expected diff is medium-large, always start with a /grill-with-docs session. Then, if planning starts to get out of hand, just say "/wayfinder let's turn this into a map".

Whether agents feel like liberation or loss depends on which joy of engineering you lean on most: making, knowing, or mattering. I've seen engineers say many conflicting things about agents: some are bubbling with excitement and for the first time ever, they feel like they're in an all-star lineup, while others mourn for the death of their craft, leaving them feeling hollow. I'm someone who works with agents every day, so I understand these sentiments. Rather than take sides, my hope is that I help you understand this in a more nuanced way. Too often we see two camps: people who love the joy of coding and people who love the joy of shipping. I think this is a gross oversimplification. Rather than asking you to pick a side, I want to ask you to reflect on the elements of your work that are the most nourishing to you. I think most of us have three sources of joy in our work, which we interweave in multiple ways: - The joy of making (or "flow"): the joy of molding things in our hand and feeling that sweet moment when it all clicks into place. - The joy of knowing: the joy of having such a strong mental model of our systems that we sense where a bug will be just by glancing at files. - The joy of mattering: the joy of knowing that you are the expert that it all relied on. Agents affect each of these elements in subtly different ways, which is why this discussion can be so complicated (we're mourning for different losses, and we're arguing from different places). The joy of making is easy to maintain. it just moves up a level. You're not shaping every line anymore. You're shaping the system or factory, in a loop fast enough that flow still happens. And when you miss the old flow, nothing stops you from writing the code yourself. The joy of knowing is especially at risk of erosion, so it's important to pay attention to it. Sometimes you'll realize you're reaching for a design you should be able to conjure automatically, but your intuition comes up with nothing. Picking an option from an agent's suggestions is a different skill than conjuring an idea in the first place. If you always pick and never conjure, the ability to conjure will erode. You can't direct something you don't even loosely understand. If you don't understand your own codebase, you're not directing an agent to do something, the agent is directing you. You matter. Obsolescence does affect you, but, you, your judgment, and especially your need for judgment, has never been more important. You get to decide what problems to solve, and you get to bring your taste and judgment to the work you do. You get to take ownership of it, all of it, whatever ships. That feeling of sadness isn't a sign of failure to adapt. Grief means that you're attached to something, and that it's natural to feel grief when you see your work changing in ways that feel unfamiliar. Lean into the things that make your work yours, and work on them more.

斯坦福大学「AI 原生软件开发」第二周课程已公开 @mihail_eric 教授带来的这门「CS146S: The Modern Software Developer」已经进行到了第二周，主题是：高级上下文工程 https://themodernsoftware.dev/ 这周课程分为两节： 高级提示技巧 + RePPIT 框架与 SDD MCP 与工具调用全解：理论、搭建与工具设计 # RePPIT 框架：五步工作流 RePPIT (Research, Propose, Plan, Implement, Test) and spec-driven development 本周的方法论主体，Mihail 用一个自己 vibe code 出来的 Mint（个人记账应用）克隆做现场演示，为它添加"按类目显示消费趋势线"的功能。关键设定是：他自己也没读过这个代码库的代码，以此模拟接手陌生代码的真实场景。 五步各自要点： 1. Research（研究） — 只记录"现状是什么"，明确禁止提出改进建议。产出一份结构化的研究文档：代码库有哪些组件、行为契约是什么、关键代码位置在哪。这份文档有双重作用：既为后续步骤播种上下文（用压缩后的表示替代喂原始代码），也帮人自己建立足够的心理模型去鉴别智能体输出的好坏。 2. Propose（提议） — 基于研究文档，让智能体给出两个相互独立的方案（附权衡分析与开放问题）。为什么是两个？讲师用 PCA 打比方：他要的是"正交的主成分"，强制智能体探索解空间的两个不同方向，更多方案收益递减，剩下的判断空间留给人。人在这两步的参与度最高：作为理解产品需求与系统约束的专家做出选择，这是当前人类杠杆最大的位置。 3. Plan（计划） — 用模板化的设计文档填充：现状、功能性与非功能性需求、设计决策与权衡、技术设计（讲师特意要求带行号，让计划接近代码级别）、测试计划、可观测性、未来考虑。其中最有价值的一节是明确圈定范围外的事项（"不要碰这些，只做范围内的工作"）这是防止智能体跑偏做无关改动的重要护栏。 4. Implement（实现） — 提示词几乎就是一句"执行计划"。真正值得注意的变化是：现代智能体（如 Cursor）会隐式地打开浏览器、造测试数据、通过 CSV 接口上传来验证自己刚写的功能，实现与测试正在合并为一步。若你的智能体没有此能力，可加 Playwright 之类的 MCP server 补足。 5. Test（测试） — 分两层。行为契约（上传 CSV → 看到趋势线）由智能体的浏览器交互直接验证；性能、可读性、安全性等超出行为契约的部分交给 agent 驱动的代码审查（实践中常见四五个专门化智能体各管一域，安全、性能、风格，由协调智能体汇总成一份报告）。 # 上下文管理的三个实操问题 这是现场问答中含金量最高的部分（依据为 LangWatch 对 2,451 个 Claude Code 会话的实证研究）： 何时压缩（compact）？ 在 25 万–45 万 token 之间执行压缩，宁可偏晚也不要太早，压缩过早的代价远高于过晚。原始研究还发现上下文利用率随窗口膨胀急剧衰减：5 万以下时利用率 47.5%，60 万以上时只剩 2.5%；而压缩后 5 步内用户纠错率会升至基线的 2.37 倍（存在"压缩后迷雾期"），所以也不能无节制地压。 何时清空重开？ 演示中讲师做了一次他自认不妥的操作：选定方案后没清上下文，导致落选方案的 token 仍留在窗口里干扰执行。他的原则是倾向于清空，把关键信息写入文件，用压缩表示开新会话。另外不要从臃肿的父会话 fork 子智能体（有人因此一天烧掉 11 亿 token），应让子智能体从干净的小上下文起步。 模型怎么分配？ 前段（研究/提议/计划）用最强的模型，因为思考集中在这里；后段（实现/测试）可以用轻量模型。这是一种直接的工程化降本手段。 # MCP 与工具调用 第二次课覆盖 MCP 的理论、搭建与工具设计。配套阅读勾勒出三个层次： 1. 协议本身：server/client/tool/transport 的基本模型——让智能体以标准化方式调用外部能力。 2. 工程实践：Cloudflare 的企业级方案直面 MCP 落地的四大痛点，供应链风险（本地 server 依赖未审计软件）、授权失控（员工各装各的）、工具 schema 洪水（52 个工具约 9,400 token，Code Mode 把它们收拢成 2 个工具约 600 token，降 94%）以及影子 MCP 检测。其"门户 + 默认拒绝 + 审计日志"的架构是 MCP 企业化的一套参考答案。 3. 工具设计的"人机工学"：为智能体设计工具不是把 REST API 直接包一层——API 的组织逻辑服务人类浏览，而智能体需要的是低歧义、可组合、schema 精简的调用面。这与上一讲的"smart tokens"一脉相承：工具定义本身也占用上下文，也是要经营的成本。 # 生态对比：三条规格驱动路线 阅读材料中的 OpenSpec 与 Superpowers 正是 RePPIT 的工业化对照物，三者共享同一思想内核（先探索、再提案、再计划、后执行、终验证），但形态不同： RePPIT：讲师自用一年半打磨的个人技能集，轻量、可自由改造 OpenSpec（约 6.8 万 star）：以 /explore → /propose → /apply → /verify → /archive 命令固化为目录化的规格产物（proposal.md、specs/、design.md、tasks.md），强调规格随工作演进且团队与智能体对齐 Superpowers（约 30 万 star）：讲师称之为"我讲的框架的强化版"，一整套可组合的技能库（头脑风暴、写计划、执行计划、代码审查、TDD、系统性调试），并强制执行"测试先行、系统化而非临时起意、证据优先于声明"的纪律。讲师特别点出：这个 30 万 star 的仓库本质就是一组 SKILL.md 技能文件，这本身就是"工作流即代码"的绝佳例证。

Should you /grill-with-docs, /wayfinder, or just one shot the code? 1. If the diff is tiny, DON'T grill /grill-with-docs saves you from situations where the agent built the wrong thing. But if the thing it's building is tiny, that cost drops to nothing. So one-shotting is absolutely fine when you're only making small changes. 2. Never start with /wayfinder A failure mode of /wayfinder is that you start grilling and the solution turns out to be way simpler than you thought. Why is this bad? Because you've created the map and tickets, and you realise you don't need them. So if the expected diff is medium-large, always start with a /grill-with-docs session. Then, if planning starts to get out of hand, just say "/wayfinder let's turn this into a map".