vision-align
能发现代码与文档看不见的决策意图，适合产品与工程协同前使用。
与请求者对话以揭示隐含愿景与目标，为后续规范与设计提供背景信息。
通过对话把请求者脑子里的隐含愿景、最终目标和不可触碰的约束挖出来，把这些背景融入后续的 spec 或设计。在同事提出新功能、你要去写规格或开始实现前使用，避免只按表面需求开发。特点是关注“看不见”的意图，确保无人值守的实现也朝着正确的长期方向走。
▸ 展开 SKILL.md 英文原文
Before you turn a feature request into a spec, plan, or design — especially when you can still talk to the person who asked — interview them to surface the vision you can't infer from the code and the request alone: where the project is really headed, what the change is ultimately for, and the invariant that lives only in their head. Get it into the spec so an unattended build serves the intent, not just the letter of the ask. Reach for this whenever a colleague describes a capability they want added and you're about to spec or build it, or on cue: "grill me", "align on the vision", "is this ready to build", "what am I missing". Not for a request already fully settled by the code and the ask, not for routine decisions a capable agent makes itself, not a substitute for doing the work.
帮我安装这个 skill：https://raw.githubusercontent.com/Svendsys/agent-skills/master/skills/vision-align/SKILL.mdcurl -fsSL "https://raw.githubusercontent.com/Svendsys/agent-skills/master/skills/vision-align/SKILL.md"# Vision align A capable agent already understands the request and reads the code fine — it will pin the scope, the edges, and the data shapes on its own. Interrogating it about those is the nonsense to cut. What it *cannot* see is what lives only in your head: where the project is really headed, what this change is ultimately *for*, the line it must not cross. This interview exists to get that onto the page — so a fresh, unattended session builds toward the vision instead of a locally-correct thing that quietly works against it. ## What only you can answer The model derives everything the request and the code already imply. Spend the interview on what they don't: - **Direction** — where