Wink Pings

After narrowing AI's room to maneuver, they opened 11,000 PRs in a month

In her AI Engineer talk, Lauren Tan said nuance tends to lead to slop. Matt Pocock broke it down into three hard rules: ban useEffect, ban 99% of explanatory comments, and use APIs to hide complexity. Commenters asked: where do the whys in comments go, and how do you block eslint-disable? The talk gave the numbers: 11,000 PRs in one month, with a rollback rate of about 0.5%.

In the AI Engineer talk "All Code Is Slop (Except Mine)," Lauren Tan (poteto) offered a judgment: nuance tends to lead to slop. Matt Pocock picked out that line when he reposted it. His point was direct: if you want to improve quality, design APIs and automated checks, and remove the ambiguity from the Agent's decision-making. Specifically, three rules: ban useEffect; ban 99% of explanatory comments, because such comments lead the Agent toward workarounds; build APIs that hide complexity where the Agent cannot see it.

Matt Pocock 推文截图

The talk video is below, at 17:31. The title is a little self-deprecating; the content is not light.

Video link: https://www.youtube.com/watch?v=O_zcc6Xl12g

This set of claims runs counter to common sense. The usual approach is to give the Agent more context, write AGENTS.md, leave explanations in the code, and tell it where the boundaries are. Lauren Tan's direction is the opposite: let the Agent make fewer judgments, and let constraints make more of them. Solomon added in a reply, "To an agent, an explanatory comment is a loophole with a citation." In other words, explanatory comments look to an Agent like a loophole with a citation. The Agent will write code around the exceptions explained by the comments, eventually making the exceptions the norm.

According to the talk, the Grok Bot team at SpaceXSI used this engineering foundation to scale AI coding agents: software architecture, lint rules, automated tests, runtime verification, and formal verification with TLA+ and Lean. Cursor agents and pstack turn these foundations into a software factory. The numbers are: 11,000 PRs in one month, with a rollback rate of about 0.5%. There is a section in the talk specifically called "Spending tokens on verification"; verification costs tokens, and that cost was not saved.

The controversy is also focused. Zero Usage asked: if explanatory comments are banned, where do you put the why behind non-obvious decisions? Matt Pocock replied that in his system, they go in ADRs. Dawnlist was not convinced, saying that comments explaining why strange things exist are exactly what the Agent needs to read. Charlie followed up: how do you distinguish comments that cause workarounds from comments that prevent bad refactors? Banning 99% sounds like guesswork. These objections do not beat around the bush, and the questions are real.

Engineering patches are more useful than slogans. One commenter pointed out that the real escape hatch from banning useEffect is // eslint-disable-next-line. Running eslint --no-inline-config in CI makes inline rule disabling ineffective; for the Agent to bypass it, it has to change the configuration, and configuration changes go into review. vinny put it more bluntly: lint errors >>> AGENTS.md line, deterministic should always be code, not prompts. Jon Neylon used design systems as an analogy: give the Agent a fixed palette and spacing, and there is no taste judgment left to get wrong. Hiding complexity in APIs and using design tokens to constrain people are the same line of thinking.

回复区截图

My observation is that this method replaces places that require aesthetic judgment with executable checks. The reason for banning useEffect is not useEffect itself. It too easily gets the Agent to do local reasoning inside dependency arrays and side effects. The comment ban does not negate comments. It moves the why into ADRs, lint, tests, and API boundaries, so machines can check it. The cost is that upfront design is heavy, rules will have false positives, and the 99% figure also looks like an empirical value. The benefit is a low rollback rate; the code submitted by the Agent looks more like factory output, with less free improvisation.

If a team wants to bring Agents into production, do not rush to add more prompts. First find the decisions that repeatedly go wrong, and see whether they can be turned into lint, types, tests, or APIs. Leave nuance for the places that need human judgment. The video and tweet are above; the comments are more interesting than the conclusion.

发布时间: 2026-11-17 12:39