# How I built a chief of staff on OpenClaw that's better than any human I've hired
**作者**: Ryan Sarver
**日期**: 2026-04-06T13:38:01.000Z
**来源**: [https://x.com/rsarver/status/2041148425366843500](https://x.com/rsarver/status/2041148425366843500)
---
I'm a VC in the middle of a fundraise, sitting on boards, helping portfolio companies, and angel investing on the side. I've worked with great human EAs and chiefs of staff over the years, so I know what high-leverage support actually looks like. When the first AI APIs came out, I tried to build an AI version of that as a product and couldn't make it work.
When OpenClaw launched I went deep immediately and haven't stopped. I have helped a number of friends set it up and each of them have asked what I have done to configure it and super power it. @ryancarson's post (link in the comments) about how he built his OpenClaw assistant was also great to see, and the response to it convinced me to finally write up what I've been building.
What I have now is more capable than any human chief of staff I've ever worked with. It never forgets a commitment, it handles the small stuff without being asked, flags the important stuff without being told, and it gets better every week. Plus it never sleeps and it never tires. There are still some bumps, but less and less each week.
If any of this is interesting, let me know. If there's enough interest I'll package the whole system up and open source it.
## What makes a great chief of staff?
Before I walk through what I've built, it's worth thinking about what a great chief of staff actually does. Not the job description, the real leverage. The best ones I've worked with filtered the noise so only the right things reached me, made sure I walked into every meeting prepared and that nothing fell through after, kept the full picture of what was in flight and flagged what was slipping, tracked relationships and knew where things stood with every important person, and created the daily and weekly rhythm that kept everything moving.
Her name is Stella. She handles all of these, and I'll walk through each one below. But the two things that make my setup genuinely different from other OpenClaw builds are the memory layer underneath it all and the continuous improvement loop that makes the system get better every week. I want to start there because they're what make everything else compound.
## Memory: the foundation
Session memory is a lie. Any assistant that treats conversation history as its working context will fail you at the most frustrating moments.
I built two layers. The first is daily notes: one markdown file per day (memory/YYYY-MM-DD.md) serving as a raw log of everything that happened. Meetings attended, decisions made, tasks added and completed, context that came up in conversation. A script called pulls from my sessions throughout the day and writes these automatically.
The second is long-term memory in MEMORY.md, curated by Stella herself. Key people, active projects, lessons learned, decisions made. She periodically synthesizes this from the daily notes, and it's what she reads on startup to orient herself on what matters right now.
Every meeting processed, every email triaged, and every task tracked feeds back into this picture continuously. Without this layer you have a capable assistant with amnesia. With it you have something closer to a person who's been working alongside you for months and never forgets anything.
I've also come to really value that all of this lives in flat markdown files rather than a database. I can open any memory file, read it, edit it if something's wrong, and understand exactly what the assistant knows. I can back the whole thing up to git and restore anything instantly. There's no abstraction layer between me and the assistant's understanding of my world, which means I trust it more and fix things faster when they're off.
Here's where the layers really come together. I'm managing a fundraise involving 100+ LP contacts across multiple countries. Stella tracks the full pipeline, keeps context on each LP and contact, and knows where every relationship stands. For first meetings, I've created a rule that she researches the fund and any recent content they or their partners have published, then preps me with what she found, how it maps to our thesis, and tailored talking points as part of the pre-meeting brief. For ongoing relationships, she knows exactly where they are in the pipeline, what was discussed and committed in our last meeting, and what the key issues are. You can't automate something as critical as a fundraise, but having this kind of structure underneath it means I'm spending my time on the conversations themselves rather than managing the process around them.
## Kaizen: the system improves itself
This might be my favorite part, and the thing that makes it feel genuinely different from any assistant I've worked with, human or AI.
Every Friday, a cron job runs research. Stella scans the OpenClaw community, checks for new patterns, looks at what other builders are doing, and saves findings to `memory/kaizen-research-YYYY-MM-DD.md`. On Sunday morning we review it together. She summarizes the week's research, surfaces the top ideas worth trying, and we talk through what to actually change.
But it's not just external research. She also learns from our daily interactions. If I keep correcting something, or if a feature creates more friction than value, that gets captured in memory and eventually surfaces as a suggestion to fix it. If a triage filter is too noisy, or a brief format isn't landing, she notices and proposes a change.
This is something a human chief of staff genuinely cannot do at scale. They can absolutely learn from working with you, but they can't simultaneously scan what hundreds of other builders are doing and cross-reference it against your system every week. That combination means the system gets measurably better on a cadence, not just when I remember to tinker with it.
The Kaizen loop also drives continuous refactoring of the system itself. I've been through multiple cycles where I build a feature, run it for a few weeks, see how it actually fits my workflow, and then clean it up or cut it entirely. First versions are always too complicated, too noisy, or solving the wrong part of the problem. A smaller system you trust will always beat a bigger one you route around. The Kaizen process makes that refactoring disciplined rather than ad hoc.
## Meeting prep and follow-through
This is where the best human EAs spend most of their time, and it's where Stella earns her keep.
Sixty minutes before any external meeting, a brief arrives via WhatsApp. It pulls prior meeting notes on the attendees from memory, checks recent email threads, and finds any open action items. For LP meetings during the fundraise, it includes pipeline stage, what deck version they've seen, their questions from last time, and what I committed to sending. I walk into every meeting more prepared than I ever was with a full-time EA.
On the other side, Stella processes every meeting through the Granola API (could be any note taker with an API). She fetches the notes, deduplicates against what's already been processed, and extracts action items. My tasks go to Todoist with proper projects and due dates. Commitments other people made get tracked in per-person markdown files, one per team member, so I can see at a glance what anyone owes or is owed. If I asked someone to do something three weeks ago and there's no update, the system knows because it's in their file.
Everything from both sides, the prep and the follow-through, feeds back into memory. So the next time I meet with that person, the cycle starts again with a richer picture.
## Task and priority management
A great chief of staff doesn't just keep a to-do list. They maintain the full picture of what's in flight and apply judgment about what actually matters today versus what can wait.
Stella keeps the comprehensive task picture in structured markdown with all the context and history. That's the source of truth. The important near-term items get synced to Todoist (again, any Task management with an API), which is where I can see my near-term tasks at a glance. I think this split is important for understanding how agents should work alongside dedicated tools. The chat interface is great for certain modes, but a focused app with a focused view is still really valuable for getting through a task list. Stella keeps both in sync, and the combination is better than either one alone.
Every evening she runs a task sweep. What's due, what's overdue, what's been sitting too long, what's coming this week. If something has rolled forward five days in a row, she flags the pattern. If there's a high-stakes meeting Tuesday and the prep isn't done, she says so. If nothing needs attention, silence.
## Stakeholder and relationship context
A great chief of staff carries a mental CRM. They know who matters, where things stand with each person, what was discussed last, and what's sensitive.
Stella does this at a scale no human could. She maintains persistent, structured context on every person, company, and project I'm working with. Relationship history, last touchpoint, open commitments, what they care about, where things stand. Every meeting processed, every email triaged, and every task tracked feeds back into this picture continuously. This is really just memory applied to relationships, but it's worth calling out because it's the piece that compounds the most visibly over time.
## Information filtering
Stella runs email and calendar triage regularly across both my personal and work Gmail accounts, surfacing what needs action and what's worth knowing while dropping everything else. But filtering is only the beginning. She auto-pulls expense receipts and routes them to quarterly tracking, generates travel itineraries from booking confirmations and flags gaps, and drafts follow-up emails as part of fundraising tasks. When a task says "follow up with [person] re: [topic]," she pulls the last thread, drafts the follow-up in my voice, and queues it for my review.
## Operational rhythm
Stella sends a morning brief at 9am and an evening wrap at 6pm, both via WhatsApp. The morning brief covers top priorities, overdue tasks, today's calendar, and anything that needs attention before I open my laptop. The evening wrap covers what happened, what stalled, and what to prep for tomorrow. If there's nothing to say, I hear nothing. This is the piece that makes it feel like working with a chief of staff rather than using a tool. She structures my day without me asking.
## Research and intelligence
A weekly curated digest of the X accounts I track and newsletters I subscribe to, organized into tiers (AI researchers, VCs, founders, operators), scored by engagement and filtered for relevance. It turns 45 minutes of scrolling into a few minutes of high value reading.
## Judgement vs predictability
One design rule that makes everything above reliable: LLMs handle judgment, and scripts handle everything else. Anything deterministic like reading files, calling APIs, sending messages, or comparing timestamps lives in Python. The LLM layer handles synthesis, prioritization, drafting, and anything where the answer requires reasoning rather than computation. When you push deterministic work through an LLM, things break in unpredictable ways and you stop trusting the system. Once you get the layer separation right, it becomes something you actually depend on.
## Where this is going
I didn't get the best assistant I've ever had by asking better questions. I got it by giving the system a better operating model. I couldn't go back, and I can only imagine where this will be in a year.
If you want to build something like this, or if you've built something similar, I want to hear about it. If there's enough interest I'll package it up and share.
## 相关链接
- [Ryan Sarver](https://x.com/rsarver)
- [@rsarver](https://x.com/rsarver)
- [469K](https://x.com/rsarver/status/2041148425366843500/analytics)
- [@ryancarson](https://x.com/@ryancarson)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [9:38 PM · Apr 6, 2026](https://x.com/rsarver/status/2041148425366843500)
- [469K Views](https://x.com/rsarver/status/2041148425366843500/analytics)
- [View quotes](https://x.com/rsarver/status/2041148425366843500/quotes)
---
*导出时间: 2026/4/7 12:40:48*
---
## 中文翻译
# 我如何在 OpenClaw 上打造了一个比我雇过的任何人都出色的幕僚长
**作者**: Ryan Sarver
**日期**: 2026-04-06T13:38:01.000Z
**来源**: [https://x.com/rsarver/status/2041148425366843500](https://x.com/rsarver/status/2041148425366843500)
---
我是一名正处于募资阶段的风险投资人，同时在多个董事会任职，帮助被投公司，还兼职做天使投资。这些年来我与很多优秀的真人行政助理和幕僚长共事过，所以我清楚高杠杆的支持到底是什么样子。当第一批 AI API 出现时，我曾尝试把它做成一个 AI 版的这类产品，但没能成功。
OpenClaw 一发布我就立刻深入钻研，再也没有停下来。我帮不少朋友搭建了它，他们每个人都问过我做了哪些配置和增强。@ryancarson 关于他如何搭建自己的 OpenClaw 助手的帖子（链接在评论区）也非常棒，看到大家的反响后，我终于决定把我一直在构建的东西写出来。
我现在拥有的这个系统，比我共事过的任何真人幕僚长都更能干。它从不忘记任何承诺，无需吩咐就处理好小事，不用提醒就标出重要事项，而且每周都在变得更好。此外它从不睡觉、从不知疲倦。虽然仍有一些小问题，但每周都在减少。
如果你对其中任何内容感兴趣，请告诉我。如果有足够多的人感兴趣，我会把整套系统打包并开源。
## 什么样才算一个出色的幕僚长？
在我介绍我构建的东西之前，值得先思考一下出色的幕僚长到底做什么。不是职位描述，而是真正的杠杆。我共事过的最优秀的那些人会过滤噪音，只让正确的事情到达我这里；确保我参加每场会议都做好准备、会后没有事情被遗漏；掌握所有在途事务的全貌并标出进度落后的；追踪人脉关系，了解与每个重要人物的关系进展到哪一步；并且建立起让一切持续运转的日常和每周节奏。
她叫 Stella。她负责所有这些，下面我会逐一介绍。但让我的配置与其他 OpenClaw 搭建真正不同的两件事，是底层的记忆层，以及让系统每周都变得更好的持续改进循环。我想先从这两点讲起，因为正是它们让其他一切都产生了复利效应。
## 记忆：基石
会话记忆是个谎言。任何把对话历史当作工作上下文的助手，都会在最让人沮丧的时刻让你失望。
我构建了两层。第一层是每日笔记：每天一个 markdown 文件（memory/YYYY-MM-DD.md），作为当天所有事情的原始日志。参加的会议、做出的决定、添加和完成的任务、对话中出现的上下文。一个脚本会在一天中从我的会话中提取内容并自动写入这些文件。
第二层是 MEMORY.md 中的长期记忆，由 Stella 本人精心维护。关键人物、进行中的项目、学到的经验、做出的决定。她会定期从每日笔记中综合整理这些内容，并在启动时读取它，以便了解当下什么是重要的。
每一场被处理的会议、每一封被分类的邮件、每一个被追踪的任务，都会持续汇入这幅图景。没有这一层，你得到的是一个患有失忆症的能力不错的助手；有了它，你得到的东西更接近一个与你共事数月且从不忘事的真人。
我也越来越珍视所有这些都存储在扁平的 markdown 文件中而非数据库里这一点。我可以打开任何一个记忆文件，阅读它，如果有问题就编辑它，并准确了解助手知道些什么。我可以把整个东西备份到 git，随时恢复任何内容。在我和助手对我世界的理解之间没有任何抽象层，这意味着我更信任它，出现偏差时修复起来也更快。
下面说说这些层真正发挥作用的场景。我正在管理一场募资，涉及多个国家的 100 多位 LP（有限合伙人）联系人。Stella 追踪整个流程，保留每位 LP 和联系人的上下文，并清楚每段关系的进展状态。对于初次会面，我设了一条规则：她会调研该基金以及他们或其合伙人近期发布的内容，然后在会前简报中为我准备好她的发现、这些内容与我们投资主题的契合点，以及量身定制的谈话要点。对于持续的关系，她清楚地知道对方在流程中的哪个阶段、我们上次会议讨论和承诺了什么，以及关键问题是什么。募资这样关键的事情无法完全自动化，但有这样的结构作支撑，意味着我可以把时间花在对话本身，而不是管理对话周边的流程上。
## 改善：系统自我改进
这可能是我最喜欢的部分，也是让它与我用过的任何助手——无论是人类的还是 AI 的——感觉真正不同的地方。
每周五，一个定时任务会运行研究。Stella 会浏览 OpenClaw 社区，查看新模式，观察其他开发者正在做什么，并把发现保存到 `memory/kaizen-research-YYYY-MM-DD.md`。周日早上我们一起审阅。她总结本周的研究，列出值得尝试的最佳想法，我们一起讨论真正要改什么。
但不仅仅是外部研究。她也从我们的日常互动中学习。如果我不断纠正某个问题，或者某个功能带来的摩擦大于价值，这些都会被记录到记忆中，并最终以修复建议的形式浮现出来。如果某个分类过滤器噪音太大，或者某种简报格式效果不好，她会注意到并提出修改建议。
这是人类幕僚长在规模化层面确实无法做到的事情。他们当然可以在与你的合作中不断学习，但他们无法同时扫描成百上千其他建设者正在做的事，并每周将其与你的系统交叉比对。这种组合意味着系统能够按固定节奏持续变得更好，而不仅仅是在我记得去调整它的时候。
Kaizen 循环还推动着系统本身持续重构。我经历过多个这样的循环：构建一个功能，运行几周，观察它实际如何融入我的工作流程，然后再清理它或完全砍掉。初版总是过于复杂、过于嘈杂，或者解决的是问题的错误部分。一个你信任的小系统，永远胜过一个你会绕开的大系统。Kaizen 流程让这种重构变得有纪律，而非临时起意。
## 会议准备与跟进
这是最优秀的人类行政助理花费最多时间的地方，也是 Stella 体现价值的地方。
在任何外部会议开始前六十分钟，一份简报会通过 WhatsApp 发来。它会从记忆中提取关于与会者的过往会议记录，检查近期的邮件往来，并找出任何未完成的行动项。在融资期间的 LP 会议上，它还会包含管线阶段、对方看过的 deck 版本、他们上次提出的问题，以及我承诺要发送的内容。我走进每一次会议的准备程度，都超过了拥有全职行政助理时的水平。
在会议之后，Stella 通过 Granola API 处理每场会议（也可以是任何带 API 的笔记工具）。她获取笔记，与已处理的内容去重，并提取行动项。我的任务会带着相应的项目和截止日期进入 Todoist。其他人做出的承诺则被记录在按个人划分的 markdown 文件中，每位团队成员一个文件，这样我就能一眼看清每个人欠了什么或被欠了什么。如果三周前我请某人做某件事而没有下文，系统会知道，因为它就在他们的文件里。
准备和跟进这两方面的所有内容，都会回流到记忆中。所以下次我再见那个人时，整个循环会带着更丰富的图景重新开始。
## 任务与优先级管理
一位优秀的幕僚长不只是维护一份待办清单。他们会掌握所有进行中事项的全貌，并对今天真正重要的事情与可以等待的事情做出判断。
Stella 用结构化的 markdown 保存完整的任务图景，包含所有上下文和历史。那是唯一的事实来源。重要的近期事项会同步到 Todoist（同样，可以是任何带 API 的任务管理工具），我可以在那里一眼看到近期的任务。我认为这个划分对于理解智能体应该如何与专门的工具协同工作很重要。聊天界面适合某些模式，但一个界面专注、视图聚焦的应用对于完成任务清单仍然非常有价值。Stella 保持两者的同步，这种组合比任何单独一个都更好。
每天晚上她会运行一次任务巡检：什么到期了，什么逾期了，什么搁置太久，这周有什么要来的。如果某件事连续五天被顺延，她会标记这个模式。如果周二有一场高风险会议而准备还没完成，她会指出来。如果没有需要关注的事项，就保持沉默。
## 利益相关者与关系上下文
一位优秀的幕僚长脑子里装着一个 CRM。他们知道谁重要、与每个人的关系进展如何、上次聊了什么，以及什么是敏感话题。
Stella 以人类无法企及的规模做到这一点。她为每一位与我打交道的人、公司和项目维护持久化、结构化的上下文。关系历史、最近一次接触、未完成的承诺、他们关心什么、进展如何。每一场被处理的会议、每一封被分拣的邮件、每一项被追踪的任务，都在持续不断地充实这幅图景。这本质上只是把记忆应用到人际关系上，但值得特别指出，因为它是随时间推移复利效应最明显的部分。
## 信息过滤
Stella 定期对我的个人和工作 Gmail 账户进行邮件和日程分拣，浮现出需要行动和值得了解的内容，同时过滤掉其他一切。但过滤只是开始。她会自动拉取消费凭证并归入季度追踪，从预订确认信息生成行程并标记缺口，并作为募资任务的一部分起草跟进邮件。当一项任务写着“就 [话题] 跟进 [某人]”时，她会调出最后一封邮件串，用我的语气起草跟进邮件，并放入队列等待我审阅。
## 运营节奏
Stella 在早上 9 点发送晨报，晚上 6 点发送晚报，都通过 WhatsApp。晨报涵盖当天的首要任务、逾期事项、今天的日程，以及在我打开笔记本电脑之前需要注意的任何事情。晚报涵盖发生了什么、什么被搁置了，以及明天需要准备什么。如果没有可说的，我就什么都收不到。正是这个部分让它感觉像是在与一位幕僚长共事，而不是在使用一个工具。她无需我开口就规划好了我的一天。
## 研究与情报
一份每周精选摘要，涵盖我关注的 X 账号和订阅的新闻通讯，按层级组织（AI 研究者、风投、创始人、运营者），按互动度打分并按相关性过滤。它把 45 分钟的刷屏变成了几分钟的高价值阅读。
## 判断力与可预测性
一条让上述一切都变得可靠的设计原则：LLM 负责判断，脚本负责其余一切。任何确定性的事情——读文件、调 API、发消息、比较时间戳——都放在 Python 里。LLM 层负责综合、排序、起草，以及任何需要推理而非计算的答案。当你把确定性的工作推给 LLM 时，事情会以不可预测的方式出错，你就会不再信任这个系统。一旦把层级划分做对了，它就会成为你真正可以依赖的东西。
## 未来的方向
我之所以得到了史上最棒的助手，不是因为问了更好的问题，而是因为给系统提供了一个更好的运行模型。我已经回不去了，我甚至无法想象一年后它会发展成什么样子。
如果你想构建类似的东西，或者你已经构建了类似的东西，我很想听听你的经验。如果有足够多的人感兴趣，我会把它打包分享出来。
## 相关链接
- [Ryan Sarver](https://x.com/rsarver)
- [@rsarver](https://x.com/rsarver)
- [469K](https://x.com/rsarver/status/2041148425366843500/analytics)
- [@ryancarson](https://x.com/@ryancarson)
- [升级到 Premium](https://x.com/i/premium_sign_up)
- [2026年4月6日 21:38](https://x.com/rsarver/status/2041148425366843500)
- [469K 次浏览](https://x.com/rsarver/status/2041148425366843500/analytics)
- [查看引用](https://x.com/rsarver/status/2041148425366843500/quotes)
---
*导出时间: 2026/4/7 12:40:48*
