当 90% 代码由 Claude 写:把验证层当正经系统来做
Claude Code 团队已让 Claude 编写和审查约 90% 的代码,Addy Osmani 的新文章讨论在这个比例下如何搭建验证层,让 Agent 拿证据证明自己改对了。文中两个做法值得借鉴:一是评判规则 REVIEW.md 固定从主分支读取,PR 无法修改评判自己的规则,新提交门禁先以 observe 模式运行,统计清楚误拦情况后再转为强制;二是一份 39 条可靠性清单,每条写明从 diff 里怎么识别、背后对应哪次真实事故,Claude 读完 diff 只执行被触发的 3~10 条。适合做 coding agent、CI 门禁与代码评审自动化的工程师细读。
The Claude Code team has Claude write — and review — roughly 90% of its code. Addy Osmani's new post is about how, at that ratio, they get the agent to produce evidence that it actually fixed what it was supposed to fix.
Two of their practices strike me as the most worth copying.
Claude Code 团队约 90% 的代码由 Claude 来写、来审。Addy Osmani 这篇新文章讲的是,到了这个比例,他们怎么让 Agent 拿出证据证明自己改对了。
我觉得最值得抄的有两个做法。
One: the change can never touch the rules that judge it. Code Review always pulls REVIEW.md from the main branch and never reads the copy inside the PR. New commit gates also start out in observe mode for a while, so the false blocks become visible before anything is enforced.
一:是改动碰不到判它的规则。Code Review 读的 REVIEW.md 永远取自主分支,不读 PR 里的版本。新的提交门禁也会先开 observe 模式跑一段,把误拦看清楚了再强制执行。
Two: a 39-item reliability checklist, where each entry spells out how to spot the problem in a diff and which incident it came from. After reading the diff, Claude only checks the 3 to 10 items that were actually triggered — nobody has to memorize the incident history.
二:是一份 39 条的可靠性清单,每条都写明从 diff 里怎么识别、背后是哪次事故。Claude 读完 diff 只查触发到的那 3~10 条,谁都不用把历史事故背下来。
Running many Claudes at once only works if the verification layer underneath is reliable enough. This post treats that layer as a proper system, which makes it worth a close read for anyone building agent systems.
Original: https://claude.dev/blog/how-the-claude-code-team-thinks-about-verification/
想同时跑很多个 Claude,前提是验证这一层足够可靠。这篇把这层当成正经系统来做,挺值得做 Agent 工程的朋友细读。
原文:https://claude.dev/blog/how-the-claude-code-team-thinks-about-verification/