# This Is How Power Users Actually Run Hermes 10x Better
**作者**: Tom
**日期**: 2026-07-27T13:12:03.000Z
**来源**: [https://x.com/tomcrawshaw01/status/2081729327541227555](https://x.com/tomcrawshaw01/status/2081729327541227555)
---
The difference between a disposable chatbot and an agent that compounds is about 30 minutes of setup, and almost nobody shows you how.
This article is based on my full video walkthrough. Watch it here: https://youtu.be/MFi3RUGzwtM
## Most people install Hermes and stop there
You got it running, connected it to Telegram, asked it a few things. And it works. You could use it exactly like that and I'd still argue it's ahead of OpenClaw and most agents out there.
But if that's all you've done, you're using a fraction of what it can do.
You re-explain your business every session. You re-teach the same process every week. You close the tab and tomorrow you're talking to a stranger again.
That's how I ran mine for weeks, until I set up the two features that change what the agent actually is: memory and skills.
## Two small files remember everything you'd otherwise repeat
Memory lives in two plain text files Hermes manages for you.
MEMORY.md is what it's picked up about your setup, your file structure, the lessons from past work.
USER.md is who you are. Your role, your preferences, your writing style, what to steer clear of. Mine even holds the persona I gave my agent, a classic northern British lad called Dave, because if I'm cloning myself in AI I may as well commit.
Both files get read at the start of every session, which is why they're deliberately small. Think sticky notes, not a journal.
When Hermes picks up a useful correction, it saves it and shows you a quick memory-updated message. You can open the files and edit anything, ask it to interview you to build a solid base, or turn on approval mode so it asks before changing anything.
## Feed it too much memory and your agent gets worse
More memory does not mean a smarter agent.
A bloated memory file makes the agent worse, because those files get injected into every session and every line costs tokens and attention.
Hermes knows this. When the files get full it doesn't pile more in, it forces itself to consolidate and keep only what matters.
Your job is the same. Review them now and then, and keep them lean.
## Skills turn a process you did once into a command you run forever
Memory helps Hermes remember information. Skills teach it how to do things.
A skill is just a repeatable process Hermes saves and reuses instead of figuring it out from scratch every time. How you answer a customer query, how you qualify a lead, how you build a weekly report.
Once a process works, you run /learn and it becomes a reusable skill. Build them from your own documents and past tasks, install community ones, or tweak the defaults Hermes ships with.
You can also chain several skills into a single command, or spin off sub-agents that split a bigger job across a small team working in parallel, instead of one agent grinding through every phase in a line.
One warning from experience: a skill is only as good as the process you teach it. If the process is messy, Hermes repeats the mess. Your V1 skill will change as you use it, and that's the point.
## One cron job makes your agent show up to work without you
Once a skill works, you can put it on a schedule with cron, and Hermes runs it automatically in the background.
This is why I run mine on a VPS. It runs 24/7, so when my laptop's closed the agent isn't. Not the only way to run Hermes, there's a desktop app now too, but 24/7 is the point.
Skills are where you stop giving Hermes more information and start giving it new capabilities it keeps using.
## It checks its own work before you ever see it
The feature that made me trust automating any of this: self-verification.
Before Hermes shows me anything, it checks its output against my own rules. If a hook doesn't carry a real number, a name, or a concrete example, it fixes it or bins it. It's judging its own work, so I don't wake up to filler.
And behind the scenes, the Curator keeps your skill library clean by quietly archiving the ones you've stopped using. If you have keepers you only need every few weeks, pin them so they never get swept.
## Wednesday, 9pm: what my agent does while I'm still at dinner
Let me show you the whole thing working, because this is where memory, skills, cron, and self-checking come together.
Every mentorship call I run gets recorded and transcribed through Granola. I've already got a written playbook for turning those calls into short talking-head reels.
The old way: me, manually mining a transcript for ideas, drafting hooks, checking each one against my recording rules.
The new way: thirty minutes after my Wednesday call ends, Hermes pulls the newest transcript, mines it with my five-point extract, drafts three talk cards in my template (a spoken hook, then Proof, Tension, and Close as bullets), verifies each one against my rules, appends the keepers to my notes file, and messages me the hooks so I can pick one from my phone.
The first cron job it set up loaded the wrong skills, and I only caught it because I asked for a summary of what it had created. Ask for the summary. Check its work once, then let it run.
## The journey timeline is the proof it's working
If you want to see the compounding with your own eyes, run /journey.
It shows a timeline of every skill Hermes has created, every memory file it's updated, and every cron job running, filling in over weeks of use.
That's the difference in one screen. Not a smarter chatbot. An agent that's measurably better this month than last month because you've been using it.
And every week you leave it unset up is another week of re-explaining yourself to a stranger.
## Set up the two files and one cron job this week
Start with the built-in memory and keep it lean. Teach it one process you actually repeat, turn it into a skill with /learn, and pin your best skills so the Curator never touches them.
Add a bigger knowledge system later, only when you actually need it. I point mine at an Obsidian vault for SOPs and client notes, and providers like Honcho exist if you want more, but the two files are the foundation.
The full build, including the two exact prompts I used, is in the video: https://youtu.be/MFi3RUGzwtM
## There's no secret setting. There's a setup
Memory gives it context. Skills give it capability. Cron makes it show up to work.
Stop repeating yourself, and Hermes starts getting better the more you use it.
## 相关链接
- [Tom](https://x.com/tomcrawshaw01)
- [@tomcrawshaw01](https://x.com/tomcrawshaw01)
- [1.8K](https://x.com/tomcrawshaw01/status/2081729327541227555/analytics)
- [https://youtu.be/MFi3RUGzwtM](https://youtu.be/MFi3RUGzwtM)
- [MEMORY.md](http://memory.md/)
- [USER.md](http://user.md/)
- [https://youtu.be/MFi3RUGzwtM](https://youtu.be/MFi3RUGzwtM)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [9:12 PM · Jul 27, 2026](https://x.com/tomcrawshaw01/status/2081729327541227555)
- [1,850 Views](https://x.com/tomcrawshaw01/status/2081729327541227555/analytics)
---
*导出时间: 2026/7/27 23:33:34*
---
## 中文翻译
# 高效用户如何将 Hermes 的性能提升 10 倍
**作者**: Tom
**日期**: 2026-07-27T13:12:03.000Z
**来源**: [https://x.com/tomcrawshaw01/status/2081729327541227555](https://x.com/tomcrawshaw01/status/2081729327541227555)
---
一次性的聊天机器人和一个能自我进化的智能体之间的差距，大约只需要 30 分钟的设置，但几乎没人教你具体怎么做。
这篇文章基于我的完整视频演示。在此观看：https://youtu.be/MFi3RUGzwtM
## 大多数人安装完 Hermes 就止步于此
你让它跑了起来，连上了 Telegram，问了几个问题。它能用。就算你只用这种方式，我依然认为它比 OpenClaw 和市面上大多数智能体都要强。
但如果这就是你所做的一切，那你只发挥了它的一小部分潜能。
你每次会话都要重新解释一遍你的业务。你每周都要重新教一遍同样的流程。你关掉标签页，明天又开始跟一个陌生人说话。
我就是这样用了几周的，直到我设置了两个能彻底改变这个智能体本质的功能：记忆和技能。
## 两个小文件记下了所有原本要重复的内容
记忆存在于 Hermes 为你管理的两个纯文本文件中。
MEMORY.md 记录了它对你的设置、文件结构以及过往工作经验中学到的内容。
USER.md 定义了你是谁。你的角色、偏好、写作风格以及需要避开的雷区。我的文件里甚至保存了我赋予智能体的设定，一个叫 Dave 的典型英国北方小伙子，因为既然要用 AI 克隆我自己，不如贯彻到底。
这两个文件会在每次会话开始时被读取，这也是为什么它们特意设计得很小。把它们当成便利贴，而不是日记本。
当 Hermes 捕捉到有用的修正时，它会将其保存并给你展示一条快速的记忆更新消息。你可以打开文件编辑任何内容，让它通过采访你来建立扎实的基础，或者开启审批模式，这样它在更改任何内容前都会先询问你。
## 喂给太多记忆，智能体反而会变笨
记忆越多并不代表智能体越聪明。
臃肿的记忆文件会让智能体表现变差，因为这些文件会被注入到每一次会话中，而每一行都在消耗算力和注意力。
Hermes 懂这个。当文件满了以后，它不会一味堆砌，而是强制自己进行整合，只保留重要的内容。
你的工作也是如此。定期检查它们，保持精简。
## 技能将你做过一次的流程变成永远可用的命令
记忆帮助 Hermes 记住信息。技能教它如何做事。
技能只是一个可重复的流程，Hermes 保存并复用它，而不是每次都从头开始琢磨。比如你如何回答客户咨询、如何筛选销售线索、如何构建周报。
一旦一个流程跑通了，你运行 `/learn`，它就变成了一项可复用的技能。你可以基于你自己的文档和过去的任务来构建它们，安装社区现成的，或者调整 Hermes 自带的默认项。
你还可以将几个技能串联成单个命令，或者分拆出子智能体，将一个大任务分发给一个小型并行团队处理，而不是让一个智能体在各个阶段按部就班地苦干。
根据经验的一个警告：技能的好坏取决于你教给它的流程。如果流程很乱，Hermes 也会重复这种混乱。你的 V1 版本技能会在使用过程中发生变化，而这正是重点所在。
## 一个定时任务让你的智能体无需你操持也能上班
一旦技能跑通了，你可以用 cron 将其加入定时任务，Hermes 就会在后台自动运行。
这就是为什么我把我的放在 VPS 上运行。它 24/7 全天候在线，所以即使我合上笔记本电脑，智能体依然在工作。这不是运行 Hermes 的唯一方式，现在也有桌面应用了，但 24/7 在线才是关键。
到了技能这一步，你就不再是单纯给 Hermes 更多信息，而是开始赋予它会持续使用的新能力。
## 在你看到结果之前，它会先检查自己的工作
让我信任这项自动化的功能是：自我验证。
Hermes 把内容展示给我看之前，会先根据我的规则检查输出。如果一个“钩子”里没有真实的数字、名字或具体的例子，它会修正或者直接丢弃。它在评判自己的工作，这样我就不会醒来看到一堆废话填充物。
而在幕后，Curator 会悄悄归档那些你不再使用的技能，从而保持技能库的整洁。如果你有一些几周才用一次的保留技能，把它们固定住，这样就不会被清理掉。
## 周三晚上 9 点：我在吃晚饭时我的智能体在做什么
让我给你展示一下整体运作，因为这里是记忆、技能、cron 和自检功能汇合的地方。
我主持的每一次导师通话都会通过 Granola 录音并转录。我已经有一份书面手册，指导如何将这些通话转化为简短的“口播”短视频。
旧的方式：我手动从转录文本中挖掘灵感，起草钩子，并根据我的录制规则逐条检查。
新的方式：周三通话结束后 30 分钟，Hermes 拉取最新的转录稿，用我的五点提取法进行挖掘，按照我的模板起草三张口播卡片（一个口播钩子，然后是证明、张力、结尾作为要点列出），根据规则逐条验证，将保留的内容追加到我的笔记文件，并通过消息把钩子发给我，让我能在手机上挑选一个。
它设置的第一个定时任务加载了错误的技能，而我之所以能发现，只是因为我要求它总结一下它创建了什么。要求它总结。检查一次它的工作，然后让它跑起来。
## 时间线是它正在发挥作用的证明
如果你想亲眼看到这种积累效应，运行 `/journey`。
它会展示一个时间线，列出 Hermes 创建的每一个技能、更新的每一个记忆文件，以及运行的每一个定时任务，这些都会随着数周的使用而不断填充。
这就是一个屏幕上的差距。不是一个更聪明的聊天机器人。而是一个因为你的使用，本月比上月 measurably 更好的智能体。
而你让它处于未设置状态的每一个星期，都是向陌生人重复解释你自己的又一个星期。
## 本周设置好两个文件和一个定时任务
从内置记忆开始，并保持精简。教它一个你真正会重复的流程，用 `/learn` 把它变成技能，并固定你最好的技能，这样 Curator 就不会动它们。
稍后再添加更大的知识系统，仅当你真正需要时。我把我自己的指向 Obsidian 笔记库用来存放 SOP 和客户备注，如果你想要更多，像 Honcho 这样的提供商也是存在的，但这两个文件才是基础。
完整的构建过程，包括我使用的两个确切提示词，都在视频里：https://youtu.be/MFi3RUGzwtM
## 没有秘密设置。只有设置。
记忆赋予它上下文。技能赋予它能力。Cron 让它准时上班。
停止重复你自己，Hermes 就会越用越好。
## 相关链接
- [Tom](https://x.com/tomcrawshaw01)
- [@tomcrawshaw01](https://x.com/tomcrawshaw01)
- [1.8K](https://x.com/tomcrawshaw01/status/2081729327541227555/analytics)
- [https://youtu.be/MFi3RUGzwtM](https://youtu.be/MFi3RUGzwtM)
- [MEMORY.md](http://memory.md/)
- [USER.md](http://user.md/)
- [https://youtu.be/MFi3RUGzwtM](https://youtu.be/MFi3RUGzwtM)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [9:12 PM · Jul 27, 2026](https://x.com/tomcrawshaw01/status/2081729327541227555)
- [1,850 Views](https://x.com/tomcrawshaw01/status/2081729327541227555/analytics)
---
*导出时间: 2026/7/27 23:33:34*