你一会儿在浏览器搜资料，一会儿回聊天框补背景，一会儿再回编辑器改几行代码。
最容易做错的，是把 Claude 当成同一种工具，以为谁分高谁就适合自己。；代价往往是如果把它们都当成同一种工具，你会在最该省事的地方继续手动搬运上下文，最后多一轮返工。；我先给一个保守判断：插件 README 第一页该是信任边界。
The cost shows up in a boring but expensive loop: research in the browser, context back in the chat box, a few edits in the editor, then back again. AI tools are starting to compete for more than code work. They are competing for the scraps of time you lose in context switching.
That is why my conservative takeaway is simple: the first page of a plugin README should be the trust boundary.
What changed my mind is that the public anthropics / claude-plugins-official repo says you should trust a plugin before you install it, update it, or use it, and that Anthropic does not control the extra tools, files, or software a plugin brings. The FAQ draws the same line: submissions get basic automated review, and Anthropic Verified is an extra layer, not a blanket guarantee. That is very different from the idea that official means safe by default.
Issue #285 made this practical. The GitHub plugin reportedly shipped with only plugin.json and .mcp.json, asked for a GITHUB_PERSONAL_ACCESS_TOKEN, and had no README or sign-in explanation. At that point, the missing document was the product, because users had to read code or guess.
Many people think they need a stronger model. Usually they need fewer windows and fewer hidden trust decisions. My scope here is narrow: the public repo, the plugins FAQ, and issue #285 as of May 24, 2026. If you build or install plugins, compare feature lists later. First ask: permissions, sign-in, dependencies, and rollback. Share this with the next person who is trying to save time with AI, not collect more windows.
真正该讨论的是：AI 工具真正开始抢的，不只是代码活，而是你来回切换的那些碎时间。