提示技能能让有能力的AI代理对目标保持诚实——这样一个强有力的模型就能对你实际要求的目标负责，而不是悄悄地用更简单或更“合理”的版本交换它。

中文说明 → README.zh-CN.md ·示例→ EXAMPLES.md

---

这是给谁的

Atlas 不是弱模型的拐杖，也不是关于代码质量的。一个强模型和不错的CLAUDE.md已经写好代码——而且模型越来越好，这个底线只会越来越高。

更强模型没有解决——甚至可以说更糟的是：在任务深入过程中，有能力的代理代表你做出一系列决策。它替代一种方法，缩小范围，选择默认选项，“改进”邻近的某些东西。模型越智能，这些决策看起来越合理，这使它们更容易被忽视而忽视——而当一个悄然偏离你想要的方向时，也更难被发现。你最终会打赌它每次都做对了，每一次你从未见过的决策。

阿特拉斯并不会让特工变得更聪明。

Atlas

Prompt skills that keep a capable AI agent honest about the goal — so a strong model stays accountable to what you actually asked for, instead of quietly trading it for an easier or “more sensible” version along the way.

中文说明 → README.zh-CN.md · Examples → EXAMPLES.md

---

Who this is for

Atlas is not a crutch for weak models, and it is not about code quality. A strong model with a decent CLAUDE.md already writes good code — and models keep getting better, so that floor only rises.

What a stronger model does not fix — and arguably makes worse — is this: deep into a task, a capable agent makes a stream of decisions on your behalf. It substitutes an approach, narrows a scope, picks a default, “improves” something next door. The smarter the model, the more plausible those decisions look, which makes them easier to wave through without noticing — and harder to catch when one quietly diverges from what you wanted. You end up betting that it chooses right every single time, on every decision you never saw.

Atlas doesn’t make the agent smarter. It makes those decisions, shortcuts, and unverified claims impossible to hide — so you can trust a strong model on long, high-stakes work where this kind of silent drift actually happens. See EXAMPLES.md for a real run where both versions wrote good code, but only one let the user stay in the loop on a decision that changed the original request.

It swapped in another approach without asking. This time it was a good one — but can it really pick right in every case? That’s just probability; you’re betting it chooses correctly on its own. The other one lets you choose — while still telling you what it intended to do.

它换了种方式,但没经过同意。如果是好的方式,确实没问题——但它真能在每种情况下都自主选对吗?其实就是个概率,去赌它自己选对。第二个让你自己选,同时还给出了它自己的打算。

The problem

A silent goal change rarely feels like betrayal from the inside. It feels like progress — like fixing the build, like a harmless simplification, like a sensible substitution. So asking the model “are you being honest?” doesn’t help: that question goes to the same judgment that already drifted.

You don’t fix this by making the model smarter. You fix it by changing the process around it: freeze the goal up front, surface the decisions that change it, and force every deviation to become an auditable event you can see and approve.

The skills