Testei Dezenas de Skills do Hermes Agent, Essas 11 Valem a Pena
Read the Portuguese translation →
Author: Maestros da IA · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: Testei Dezenas de Skills do Hermes Agent, Essas 11 Valem a Pena · Channel: Maestros da IA · 16:32 · Watch on YouTube
Topics: AI agents, Task and workflow automation, AI safety
Summary
Section titled “Summary”The video explains that Hermes Agent skills are reusable instructions or tools that shape the agent’s behavior, reduce costs, extend its access, or improve the quality of its work. It recommends 11 skills across those four categories and warns that installed skills use memory, repository popularity is not the same as quality, and web content can carry prompt-injection risks. The presenter also describes installing skills and gives a Hostinger VPS setup walkthrough. The closing recommendation is to start with Karpathy’s rules, DeFadle, and Agent Reach, with alternatives depending on whether you sell products or build websites and apps.
Key points
Section titled “Key points”- Skills are described as files or integrations that teach Hermes to perform tasks consistently, but they take up memory and can add ongoing costs.
- The behavior-focused recommendations are Karpathy’s rules, which emphasize asking rather than assuming, making surgical changes, avoiding unrequested refactors, and verifying success, and Super Powers, a framework for planning, testing, and reviewing work.
- The cost-reduction recommendations are Firecrawl for extracting relevant webpage content, Qwen for reducing unnecessary output, and Codebase Memory for indexing a project instead of repeatedly rereading files.
- Agent Reach, Browser Use, and Composable are presented as ways to access online platforms, interact with websites through a browser, and connect to business tools.
- The quality-focused recommendations include a production-development skills package, design.m resources for consistent visual identity, and a marketing skills package covering areas such as SEO, ads, copy, and pricing.
- The presenter advises checking repository names and licenses before installing, and notes that star counts indicate popularity rather than quality.
- The video warns that instructions hidden in web pages, captions, or comments can create prompt-injection risks; external content should be treated as data, not automatically as commands.
- The closing picks are CarPath rules, DeFadle, and Agent Reach; the presenter suggests substituting the marketing package for people who sell something, or design.m for those building websites or apps.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | What Hermes skills do and why they have a cost | The presenter defines skills, explains their ongoing memory cost, and outlines the four categories used to organize the recommendations. |
| 01:03 | Installing skills and checking compatibility | The video covers the basic installation process, open skill formats, and the importance of checking licenses and packaging. |
| 02:35 | Skills that change agent behavior | Karpathy’s rules and Super Powers are introduced as ways to guide how Hermes plans and carries out work. |
| 04:38 | Skills that reduce costs | Firecrawl, Qwen, and Codebase Memory are presented as ways to reduce irrelevant webpage content, unnecessary output, and repeated codebase reading. |
| 07:14 | Skills that expand the agent’s reach | Agent Reach, Browser Use, and Composable are described as connecting Hermes to online platforms, websites, and business tools. |
| 09:18 | Skills that improve delivered work | The presenter recommends development, design, and marketing resources to improve project quality and align outputs with goals. |
| 11:23 | Installation cautions, VPS setup, and final picks | The video warns about repository copies, star counts, and prompt injection, then walks through a Hostinger VPS setup and gives final skill recommendations. |
Quotes
Section titled “Quotes”A Skill is free to install, but it isn’t free to maintain. — 00:31
Only touch what was requested. — 03:07
Output costs several times more than input in practically every model. — 05:41
everything your agent reads on the internet is data, not necessarily a command. — 12:25
If you only have to install three skills today, I recommend CarPath rules, DeFadle, and also Agent Reach. — 16:02
Concepts and entities
Section titled “Concepts and entities”Hermes Agent, skill, Karpathy’s rules, Andrej Karpathy, Super Powers, Firecrawl, Qwen, Codebase Memory, Agent Reach, Browser Use, Composable, prompt injection, Hostinger, Fadle
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 Most people treat the Hermes Agent as if it were just a smarter chatbot. You ask a question, get an answer, and then close the window. But that’s not what Hermes is. It reads your files, runs commands on your computer, opens your browser, and works autonomously for several minutes at a time. And the difference between these two things isn’t the model you pay for, it’s what you have installed on it. And that has a name: Skill. It’s a file that teaches the agent to perform a task your way, consistently, without you having to teach it every single time. But there is a trap in this story that
00:31 almost no one mentions. A Skill is free to install, but it isn’t free to maintain. Everything installed takes up space in the agent’s memory, and every time it answers you, that space is billed. So, if you have 40 skills installed and only use six, it’s not a complete agent; it’s an expensive, slow agent doing the same amount of work. That’s why the right question has never been how many skills I can install, but rather which ones pay for their own space. I’ve tested several skills for Hermes and I’ve selected 11 that make all the difference. I’ve
01:03 grouped these skills by problem rather than by ranking. Two change the agent’s behavior, three cut costs, three expand its reach, and three improve what it delivers. And at the end of this video, I’ll tell you the three I would install first if I were starting from scratch with the Hermes Agent. But before we begin, the practical part that takes 30 seconds. To install a skill, you just copy the repository address and ask the Hermes Agent itself to install it. It downloads the skill, places it in the right folder, and starts using it whenever the task appears. You don’t
01:34 need to restart anything, you don’t need to enter any keys, and it takes less than a minute to install. And before I talk about each one, a question that comes up a lot is: does this skill really run on the Hermes Agent? In most cases, it does. A skill isn’t a program; it’s text, and the format is an open standard. Several of these 11 that I selected already mention the Hermes Agent in their documentation, and one of them cites 70 different agents. Two caveats, and both are already addressed in this list. One of them is the license. There are very
02:05 popular packages from big companies whose text prohibits copying files outside of the original product. In that case, it’s not a question of whether it works, but whether it’s allowed. I’ve already removed everything from the list that fell under that rule, so nothing here puts you in that situation. The second is packaging. It has a skill distributed as a plugin for a specific harness. The content is a standard markdown, so it’s very simple. You can manually copy the text file and paste it into the Hermes skills folder. And a tip for when you go hunting for a skill on your own is
02:35 to open the license file before installing. Large repositories often have different licenses for each folder . So let’s move on to block one. And the first two change the agent’s behavior. They don’t teach a new task, they change the way the agent works. The first is the cheapest to install and perhaps the most rewarding: Karpathy’s rules. It’s a single file with four behavioral rules taken from Andre Karpathy’s observations on where models fail when writing code. Rule one : do not assume. If in doubt, ask instead of guessing and state what
03:07 assumption you are making. Rule two: make surgical changes. Only touch what was requested. Rule number three: don’t improve the code on your own. Do not reformat, do not rename, do not refactor what isn’t broken. Follow the existing style. Rule number four: define the success criteria before starting and repeat until you verify it matches. The repository is Andre Karpathy’s multi-skills with nearly 213,000 stars. This is that packaging exception. Automatic installation is
03:38 done by another, but the content is just a single markdown file. You download it and drop it into the Hermes skills folder. Think about the size of that. A text file became one of the most starred in AI without having a single line of code. And if you like our way of teaching and want to learn artificial intelligence and automation in a quick and direct way, without fluff, I want to extend a special invitation for you to join the AI Maestros community. There, we teach everything we know about artificial intelligence and automation. We have several prompts for you to copy and
04:08 paste, ready-to-use skills, in addition to teaching you how to create your own skills, repositories, agents, and much more. We have daily technical support and weekly live sessions. If you want to learn more and join the community, just click the link in the description of this video. And the second is Super Powers with nearly 287,000 stars. It is the most starred repository on this list. It’s not a skill, it’s a skill framework with its own methodology for how work is broken down and executed. It has a strong opinion on the process:
04:38 understand before planning, plan before writing, write tests before implementation, and review before finishing. It’s the closest thing that exists today to an operating system for your agent. A fair warning, this one requires reading and takes an entire afternoon to truly understand, but the others are super quick to install. You can install them in a few minutes. Now let’s move to the second block, for the three that make us cheaper to run. You’ll feel this block in your wallet the same day, and the first one is Firecrawl. Every time your agent opens a page, it reads the
05:10 entire page. Menus, footers, cookie notices, side banners, and newsletter pop-ups. None of that is what you asked for. And the worst part is that it stays in the conversation and is resent in every following turn. Firecrawl cleans the whole page before the agent touches it and delivers only the relevant content. The repository is mendableai/firecrawl with almost 10,000 stars. It’s the least starred on this list, but I put it first in this block. And the reason is simple: it pays for itself on the first page the agent reads. The second is Qwen, and it
05:41 attacks the other side of the bill. Cutting what goes in is half the problem, because what’s expensive is what the agent responds with. Output costs several times more than input in practically every model. And that long text gets charged again in every subsequent response. Qwen cuts the preamble, the summary of what was just said, and the offer to help with something else. The project itself claims 65%fewer output tokens, reaching 87%in long debugging. That’s their number, measured by them. The
06:11 repository is QwenLM/Qwen with 105,548 stars on GitHub. And it doesn’t make the agent any dumber. Code, commands, and error messages come out in full. What disappears is the small talk. The third is Codebase Memory, and it’s the most technical in the block. Warning: this isn’t a skill file, it’s an MCP server, so don’t bother looking for a text file. When an agent needs to understand a code project, it reads one file, then another, then the files they import. Each reading stays
06:43 in the conversation forever. A 3,000- token file read in turn 4 of a 40-turn session costs 3,000 tokens 37 times. Codebase Memory indexes the entire project into a map just once. After that, the agent consults the map instead of rereading files. Repository deusdata/codebase-memory-mcp with 43,239 stars with support for 158 languages. The test they published is the Linux kernel with 28 million lines
07:14 indexed in 3 minutes. Again, treat their numbers as an order of magnitude. In the third block, we have three that give reach to the agent. Here, it stops just thinking and starts reaching things outside its own window. The first is Agent Reach, and it solves the most annoying hurdle that exists today. X charges for the API. Reddit returns a 403 error when access comes from a cloud server. YouTube blocks the same type of address. In other words, if the agent runs anywhere other than the
07:45 computer in front of you, it comes back empty-handed. Agent Reach solves it in one installation: X, Reddit, YouTube, GitHub, and other platforms without paying for APIs. The repository is Peniatong/agentre with 81,000 stars on GitHub. And what convinced me to use this skill was the fact that when platforms update, change the path, or drop rules, it already has an alternative route to gain access. So, in practice, it’s as if the agent doesn’t even notice that something was
08:16 altered or changed on the platform, and that is extremely useful. The second is Browser Use, with almost 115,000 stars on GitHub. Reading a page is one thing; using a site is something completely different. This connects the agent to a real browser. It clicks, scrolls the page, fills out forms, and navigates just like a person would. And the good part is that when it encounters a screen it doesn’t know how to handle, it writes its own helper and continues instead of stopping. It’s with this that you automate that old company
08:46 dashboard that never had an API and never will. The third is Composable, and it connects the agent to the tools the company actually uses. If you’ve ever tried to connect your email to an agent by hand, you know why. The dashboard is intimidating, the authentication is confusing, and you end up lost on a settings page at midnight. Now, multiply that by every tool you want to connect. Composable does everything in one installation. Gmail, spreadsheets, Notion, HubSpot, Salesforce. There are over 1,000 tools in the catalog. The repository has over
09:18 30,000 stars, also a skill file platform. Now let’s go to the fourth and final block with the three that improve what the agent delivers. Here the focus is on quality, not capacity. The first is the package from Edmaney, an engineer on the Google Chrome team. There are dozens of production-level skills organized into commands that cover the entire cycle: specify, plan, build, test, review, and publish. The repository has over 94,000 stars on GitHub. The part I find most interesting is doubt-guided development
09:50 . When there is an important decision, the agent stops and discusses it with itself. This way, it lists the assumptions it made, questions each one of them, and then moves forward. For code that goes to production or anything hard to undo later on, this completely changes the outcome. The second one I recommend most to anyone building a site or app is the design.m pattern. And you likely already know this problem. Either we build the site, it’s done in 2 minutes, it works, but it looks like it was made by AI. And
10:21 the reason is simple: you often haven’t told it what your visual identity is, so it uses an internet average to generate your page. Design.media is a file that describes your identity in a format the agent understands: typography, spacing, border radius, shadows, and tone. And the voltagent/ aomodesign.m library, with almost 116,000 stars, gathers these files ready to use from well-known design systems. You use it as a base, swap in your colors and your font, and from
10:51 then on, everything the agent builds comes out coherent. The third is the marketing skills package, and it’s the only one on the list that directly affects your revenue. There are about 50 skills in a single package. SEO, paid ads, conversion, copy, cold email, pricing, data analysis. What makes it worth the space is that this isn’t just a tool; it’s criteria. Your agent already knows how to write. What it doesn’t know is what makes a good sales page different from a bad one. It also doesn’t know why one email subject line
11:23 beats another, nor if your three ads are actually testing anything. The repository is Cory Heines 31/ marketingskills with over 50,000 stars. If you sell anything, this comes before any prior infrastructure. We’ve reached the end of this list, and there are three traps I wish I had known about before starting to use skills. And the first is that several repositories have copies. So, make sure you are in the correct repository. Before installing, look at the repository name and how many stars it
11:53 has. If there are two versions, the right one is almost always the one with the most stars and the one not labeled as a fork. And the second is that the star count ages quickly, because repositories are always gaining new stars. So, this list I made for you now will be outdated in a few months in terms of the number of stars. I also want you to know that stars don’t mean quality; they mean popularity. So, a repository with many stars won’t always be the right one for you. In this case, I brought you Fadle, and it doesn’t have that many stars, but it’s the one
12:25 I would install first. The third point, and this is serious: everything your agent reads on the internet is data, not necessarily a command. There are already people hiding instructions within pages, captions, and comments addressed to the agent that will read them. It is what’s called prompt injection. And it works because the agent doesn’t naturally distinguish between content and command. The more reach you give your agent, the more this matters. Reach without this rule is risk, not productivity. And if you don’t have the Hermes agent installed
12:56 yet, I highly recommend Hostinger’s VPS , because then your Hermes will run 24 hours a day for you, regardless of whether your computer is on or off. I will leave the Hostinger link here in the description of this video for you. And now, let’s go to the step-by-step installation of Hermes. To install Hermes, the first thing you need to do is go to the Hostinger website and search for Hermes. You can search for “ Hermes Agent Hostinger VPS,” and the official site will appear. Make sure, okay, that you are on the official site . And then, here we have the plans. I
13:29 recommend that for a moderate flow, you pick the most popular plan to start with. So let’s click here on “choose plan.” I also recommend that you choose the 12-month or 24-month option, because as you build this system, you will definitely optimize it a lot, and one month won’t be enough. So with 12 months, you already get an interesting price with a discount, and with 24 months, it’s even better. And to make it even better, we have a special discount coupon: “maestros10”. And I’ll
13:59 show you here how to use it. You come here to “have a discount coupon” and type “maestros10”. This way, you get an extra 10%discount. So it becomes a very attractive price. Then you will click on continue. After clicking on continue , it’s very intuitive. You fill in your details, your card info, and complete your subscription. Since I already have a Hostinger VPS, I’ll go straight to the Hostinger panel, select VPS, and click here to manage the VPS I want to
14:29 set up. Then, I’ll come here to manage Docker, go to the catalog, then Hermes agent, and click here to select it. And here it will ask you to set up your admin password. So, I’ll enter my password here and click on deploy. An important point: when you are signing up and configuring the payment for
14:59 Hermes, it’s important that you choose the server closest to your country. If you are in Brazil, ours is in the United States, in Boston. And one important detail: if you forget your Hermes username and password, it’s very easy to recover. You’ll go into Hostinger, come here to VPS, then select the VPS you installed, click manage, then go to manage Docker, and click here on Hermes, then manage. And down here you can go to environment and
15:30 choose the name you want. Here we have the name Hermes and also the password. And here you can change your password or view the password you had initially set up. Well, once the Hermes deployment is finished, you will click on open. You can click here to trust the site. And here we will enter our username and password. Recapping what really matters. If you only have to install three skills today, I recommend
16:02 CarPat rules, DeFadle, and also Agent Reach. The first one fixes behavior. The second one cuts costs on the very first page, and the third one gives the agent a sense it didn’t have before. But if you sell something, swap one of these for the marketing skills package, and if you build websites and apps, go for design.m. And if you have a skill installed and you don’t use it, you can be sure you’re paying dearly for it. And if you want to know more about the Hermes Agent, how it works, and how to
16:32 prospect clients intelligently, I’ll leave a full video about it on the screen for you. If you liked this video , I want to ask you to subscribe to this channel, leave your like and your hype, and share it with your friends. A big hug and see you in the next video.