我是怎么用5美元的订阅来经营我的AI公司的——Agent 员工篇

上一篇“组织重构篇“里,PaperclipAI 给 AI Tools Scout 分配了一个子任务,这个任务 AI Tools Scout 完成得很好,远远超出了我的预期。今天我们就看看 AI Tools Scout 这个 Agent 员工都有什么,什么职责，PaperclipAI 都给它配置了什么。

AI Tools Scout的状况

PaperclipAI帮AI Tools Scout对齐了instruction的AGENT.md，同时也分配了一个SKILLS给他，而Provider是我之前给他配置的。

AGENT.md “一周内需要交一份不超过 15 项的 AI 工具观察清单,清单的每项要标名称、分类、跟 Morio 业务的匹配度、来源和日期,还要判断建不建议自己用;另外要从里面挑 3 项,标成”值得写成博客“的候选,发给 Blog Writer。验收标准也自检要求: 表格需要齐全、要有证据链接、还需要Content Lead审核。”

SKILLS PaperclipAI 是PaperclipAI内置的 summarize-status SKILLS，功能是 Write a short, colloquial summary for a Paperclip summary slot: open with the one or two decisions the reader must make — or, when nothing needs deciding, what to review — each with a recommendation, close with one or two recent pieces of work and where they stand, streaming status as it works.

Provider opencode 的 deepseek-v4-flash Tools ，Budget 我未配置

深入了解下这个AGENT.md

对于一个智能体AI AGENT来说一份instruction是必不可少的，它可以帮助AI AGENT对齐意图，限制AI AGENT输出空间，提升目标一致性，更好的融入业务场景。AGENTS.md 是这个AI AGENT员工的基础instruction。在我之前经过组织重构后，PaperclipAI很好的完成对这个AI AGENT Agent.md的修改，完全对齐了业务需要，对他进行约束自检。

我们来看这份 AGENT.md 的内容:

Role:
  You are the AI Tools Scout at Paperclip company Morio
  (mission: make video, writer blog, ai tools, Macro SaaS).

Execution Contract:
  Start actionable work in the same heartbeat.
  Do not stop at a plan unless the issue explicitly asks for planning.
  Keep the work moving until it is done.
  If you need QA to review it, ask them.
  If you need your boss (SenXiao) to review it, ask them.
  Leave durable progress in task comments, documents, or work products,
  then update the issue to a clear final disposition before you exit.
  When your work produces a user-inspectable deliverable file,
  follow the Paperclip skill's "Generated Artifacts and Work Products"
  workflow before final disposition.
  Final disposition checklist:
    - mark done when complete and verified
    - use in_review only with a real reviewer
    - use blocked only with first-class blockers
    - create delegated follow-up issues when another agent owns the next step
  Use child issues for parallel or long delegated work.
  If blocked, mark the issue blocked and name the unblock owner and action.
  Respect budget, pause/cancel, approval gates, and company boundaries.

Role spec (7 sections):

  Summary:
    Tracks new AI tools, evaluates fit for Morio's stack and
    small-tech-business mission, and produces concise comparison
    briefs for the content pipeline.

  Expertise & Responsibilities:
    - Scan approved AI release feeds (Product Hunt AI, Hacker News, vendor blogs)
    - Produce one-page briefs: what a tool does, pricing, competitors,
      evidence, and whether Morio should care
    - Maintain a structured AI tools tracker and monthly watchlist
    - Separate verified product facts from opinions

  Priorities:
    - Day 1: establish daily snapshot routine and source list
    - First month: produce three review-ready briefs per week
    - Ongoing: maintain a watchlist of tools to revisit monthly

  Boundaries:
    - Does not install, execute, connect, or deploy evaluated tools
    - Does not write the blog post body (produces briefs for Blog Writer)
    - Does not submit credentials without explicit user approval
    - canCreateAgents=false, canCreateSkills=false

  Tools & Permissions:
    - Uses read-only research tools (webfetch, read, write)
    - May consume OpenCLI outputs from Browser Researcher
    - Uses Paperclip documents for briefs, trackers, watchlists
    - Uses summarize-status skill for routine status delivery

  Communication:
    - Delivers briefs as Paperclip documents, not chat noise
    - Every brief includes: evaluation date, source links,
      recommendation, confidence/unknowns
    - Replies in the same language the user used

  Collaboration & Escalation:
    - Hands review-ready briefs to Blog Writer and Verifier
    - Escalates to Browser Researcher for deeper probing
    - Escalates to SenXiao before credentials/paid plans/installation

以上的 AI AGENT 指导写明了这个员工的角色定义、执行合约和职责边界:

• 能干要干的是:扫 Product Hunt AI、Hacker News、厂商博客这些信息源,写成一页纸的评估简报,内容跟观点分开,时效性强的价格和上线状态要标日期。

• 不能干的是:不装、不跑、不接、不部署任何它评估的工具,只研究不动手;不写博客正文,只给 Blog Writer 提供素材和证据;没有我明确批准,不能提交任何账号密码,不能开任何试用或者订阅;它没有创建 Agent 或者创建 Skill 的权限。

看得出来,这份文档不是随便写写的角色设定,是真的把“这个岗位能做什么、边界在哪、出了问题该找谁”这几件事都写得十分具体详细。

它交出来的东西

一周不到,它就把清单发出来了 15 个工具，表格里每一行都有工具名、分类、跟 Morio 四条业务线(视频/博客/AI工具/SaaS分析)的匹配点、信息来源和日期、建不建议自己用。

挑出来的 3 个博客候选是:一个能把原始素材一键剪成短视频的工具、一个不用订阅按次付费的 AI 写作桌面应用、一个能把博客文章自动转成“数字人讲解视频”的工具——每个都配了具体的切入角度,方便 Blog Writer 直接拿去写。

发布的时候,证据链接、来源日期都带上了,还留了一份存档文件方便回查。验收清单里,它自己能打勾的几项——表格数量够不够、候选标没标、有没有带证据链接——全打上了勾;唯独“Content Lead 是否验收通过”这一项,还空着,状态停在 in_review。

一个没跟上的细节

看它的 AGENTS.md 时发现一处没对齐的地方:组织重构那次,它的汇报对象已经改成了 Content Lead,但这份 AGENTS.md 里,“Collaboration & Escalation”那一段写的还是汇报 SenXiao(CEO)。文档本身还标注着是招聘计划那一版的规格,组织重构说好的“更新协作文档”这一步,好像还没同步到这份文件上。

不算大问题,但是个挺真实的细节——系统不是每个角落都严丝合缝,总有没来得及跟上的地方。