"处理我的邮件。"

"Handle my emails." Good prompt: "Done means every morning at 9:00 AM the inbox is empty, replies are drafted in my voice, and VIP escalations are flagged to the top." 👤 2. Give it an Identity. An LLM out of the box knows a little about everything, making it generic. Giving it an identity focuses its genius. You need 3 simple plain-text configuration files to ground it: The Soul File: Defines its personality, values, quirks, and exact communication voice. The Identity File: Its name, specific DNA, core role, and strict boundary parameters (e.g., "Inbox only. Never touch my calendar."). The User File: Massive context about you—your goals, your business ecosystem, and who the agent is serving. 🛠️ 3. Equip It. Context is your agent's ultimate moat. To let it work effectively, you must map your workflows. The best way? Reverse prompt the source. Instead of writing a manual playbook, connect your agent to past logs (like your last 50 sent emails) and prompt it: "Study my tone, sentence length, and best practices. Generate a style guide capturing my voice, and draft the next unread email as a test." 🔎 4. Narrow the Scope Avoid the "Mega-Agent" trap. Do not build one agent to handle marketing, coding, and admin—it leads to massive context rot. You need one specialized agent per lane. For complex workflows, build a Manager Agent whose sole job is to orchestrate, delegate, and monitor specialized sub-agents underneath it. 🤝 5. Trust in Stages. You don’t hand over the keys to the car on day one. You build trust systematically: Guardrails: Define hard code boundaries inside the identity files. Review: Set it to generate drafts/actions for approval first. Loosen the leash: Allow autonomous execution on low-risk tasks (like forwarding sorting logs). Heartbeat: Turn on the recurring timer loop so it executes solo every 15 minutes. Where to start?Look for tasks that fit the Rules of R: Is it Repetitive? Is it Rules-based? Does automating it generate a massive Return on Time? Stop babysitting chatbots. Start orchestrating your digital workforce.

"处理我的邮件。" 好的提示词："完成意味着每天早上 9:00 收件箱为空,回复以我的语气起草,VIP 升级事项置顶标记。" 👤 2. 赋予它身份。LLM 开箱即用,对什么都略知一二,使其显得泛泛。赋予它身份能让它的才能聚焦。你需要 3 个简单的纯文本配置文件来将其落地: 灵魂文件(Soul File):定义它的个性、价值观、怪癖以及精确的沟通语气。 身份文件(Identity File):它的名称、具体基因、核心角色以及严格的边界参数(例如:"仅处理收件箱。绝不动我的日历。")。 用户文件(User File):关于你的海量背景信息——你的目标、你的业务生态系统,以及该代理服务对象。 🛠️ 3. 为它装备。上下文是代理的终极护城河。要让它高效工作,你必须梳理你的工作流。最佳方式?对源头进行反向提示。与其写一本手动操作手册,不如将你的代理连接到过去的日志(例如你最近发送的 50 封邮件),并提示它:"研究我的语气、句子长度和最佳实践。生成一份捕捉我语气的风格指南,并将下一封未读邮件作为测试进行起草。" 🔎 4. 收窄范围。避开"Mega-Agent"(超级代理)陷阱。不要构建一个代理来处理营销、编码和行政事务——这会导致严重的上下文腐烂。每个赛道你需要一个专业代理。对于复杂工作流,构建一个管理代理(Manager Agent),其唯一职责是编排、委派和监控其下的各个专业子代理。 🤝 5. 分阶段建立信任。你不会在第一天就把车钥匙交出去。你系统地建立信任: 护栏(Guardrails):在身份文件中定义硬性代码边界。 审核(Review):将其设置为先生成草稿/操作以待批准。 松绑(Loosen the leash):允许在低风险任务(如转发、排序日志)上自主执行。 心跳(Heartbeat):开启循环定时器,使其每 15 分钟自动独立执行一次。 从哪里开始?寻找符合 R 法则的任务:它是否是重复性的(Repetitive)?它是否是规则驱动的(Rules-based)?自动化它是否能产生巨大的时间回报(Return on Time)? 别再像保姆一样照看聊天机器人。开始编排你的数字员工队伍。