Dan Luu 写系统问题时，习惯先测量、再对照、最后才下结论。把「Some thoughts on writing」放到智能体、评测和线上系统里，真正要问的是：默认做法会不会系统性失败。下面用中文整理成可执行的工程笔记，去掉原站导航和无关链接。

问题怎么暴露

I see a lot of essays framed as writing advice which are actually thinly veiled descriptions of how someone writes that basically say "you should write how I write", e.g., people who write short posts say that you should write short posts . As with technical topics, 一个稳妥判断是 a lot of different things can work and what's really important is that you find a style that's suitable to you and the context you operate in . Copying what's worked for someone else is

We'll start by looking at how much variety there's been in what's worked 1 for people, come back to what makes it so hard to copy someone else's style, and then discuss what I try to do in my writing.

测量时容易踩的坑

If I look at the most read programming blogs in my extended social circles 2 from 2000 to 2017 3 , it's been Joel Spolsky, Paul Graham, Steve Yegge, and Julia Evans (if you're not familiar with these writers, see the appendix for excerpts that 一个稳妥判断是 are representative of their styles ). Everyone on this list has a different style in the following dimensions (as well as others):

To pick a simple one to quantify, length, Julia Evans and I both started blogging in 2013 (she has one post from 2012, but she's told me that she considers her blog to have started in earnest when she was at RC , in September 2013, the same month I started blogging). Over the years, we've compared notes a number of times and, until I paused blogging at the end of 2017, we had a similar word count on our blogs even though she was writing roughly one order o

对照之后能下的结论

To look at a few aspects that are difficult to quantify, consider this passage from Paul Graham, which is typical of his style:

What nerds like is the kind of town where people walk around smiling. This excludes LA, where no one walks at all, and also New York, where people walk, but not smiling. When I was in grad school in Boston, a friend came to visit from New York. On the subway back from the airport she asked "Why is everyone smiling?" I looked and they weren't smiling. They just looked like they were compared to the facial expressions she was used to.

落到生产里的动作

If you've lived in New York, you know where these facial expressions come from. It's the kind of place where your mind may be excited, but your body knows it's having a bad time. People don't so much enjoy living there as endure it for the sake of the excitement. And if you like certain kinds of excitement, New York is incomparable. It's a hub of glamour, a magnet for all the shorter half-life isotopes of style and fame.

Nerds don't care about glamour, so to them the appeal of New York is a mystery.

值得单独记下的观察

• Level of technical detail
• Amount of supporting evidence
• Explain a technical topic that a lot of people don't seem to understand at a level that's accessible to almost any professional programmer Examples: branch prediction , malloc , cache partitioning
• Make a case for a minority opinion (or one that was a minority opinion at the time, anyway): Examples: files are difficult to use , public tech companies can pay very well , monorepos aren't stupid
• Discuss phenomena I think are interesting: Examples: funny discontinuities , the difficulty of knowledge transfer , normalization of deviance
• Low up-front investment, with as little friction as possible, maybe increasing investment over time if I continue blogging
• Improve writing technique/ability with each post without worrying too much about writing quality any specific post
• Only publish when I have something I feel is worth publishing

落地时建议先做的 5 件事

1. 用自己的真实负载测，而不是只用公开榜或厂商数字。
2. 把评测设计成能抓到失败模式：平均分好看但尾部崩溃，仍然算失败。
3. 智能体默认不会好好用测试；要写进流程，而不是写在口头规范里。
4. 性能和正确性都要有基线，改模型或改语言前后必须能对比。
5. 结论写成可回滚的决策：哪一版配置、哪一版评测集、谁签字。

和智能体产品怎么接

龙虾PRO做 OpenClaw 落地时，同样吃「先测量再扩面」这条纪律：技能、数字员工和网关都要有可复现评测，而不是只看一次演示通过。

本文侧重全链路风控方法论。落地时请用自身业务单据做回放验证，不要把示例阈值直接当生产策略。 相关：风控体检 · 方案资源