How To Build A Hermes Agent Skill In 15 Minutes transcript

A skill file you write yourself in 20 minutes will beat almost every skill you will ever download. I tested 55 of the most popular skills online and 27 of them cannot tell your agent when to run. Yet, they load on every single message you send your agent, cost you a ton of your money, and the worst part is they give you nothing in return. But, the five skills I wrote myself are the only ones I have never turned off and also the ones doing the most work for me in my business.

So, in this video, we are building one skill together in five simple steps. By the end of this video, you will never have to explain that job to your agent ever again, and you will have an agent that feels like a full-time employee for free. Let's get started. The skill we are building together in this video will audit your other skills. But, you can use everything we are about to go through together for any skill you want to build yourself for absolutely any use case you need.

But, today's skill will read every skill we have installed. It will have to check whether their descriptions actually say when to use it. It will measure how big each one is. Then, it will hand us an audit table. I picked the skill on purpose for this video, well, because it felt very fitting for the topic of this very video. I'll even give you the skill we built together by the end of this video so you can use it your And the first thing we need to get right when building a skill is not the file.

It is where the content comes from. Before anything else, you need to start by picking your job. Not mine. Pick what skill you want your agent to do for you. A pro tip here, try thinking of things you've asked your agent for multiple times before. And if you can't think anything right away, just ask your agent, "What are the things that I ask you for repeatedly that you think we should turn into a skill?" And whatever that answer is for you, that is the skill you should build first.

Feel free to build your own skill as you're watching this video along together. Now that we can start, here's how you should think about skill building. A skill is worth building when it holds something your agent can't work out on its own. Your way of doing a job, your corrections, your specific mess. So, there are two honest places to get that. Number one, you either do the job by hand once and write down every correction you made along the way, or you point it at work you have already done and let it read the corrections out of that.

Here is the sentence I would actually type out. "Go back through the last three times we did this, insert your skill, and write down every single thing I corrected you on. Do not summarize it. I want the actual corrections. Those corrections are the most valuable thing in the whole file. Every one of them is a mistake you never have to catch again. If you skip this and you rewrite the same note every week for the rest of your life." Another pro tip here, you can literally tell your agent, "I want to create a skill called XYZ, whatever your skill name is.

Every time I run it, you need to do ABC. Let's do it together first, step by step, so you can learn how to do it exactly as I want it done." Look, this step is like 80% of what makes a good skill, and it is the one everybody skips because it feels like it's a lot of work. And once you have the raw material, there is one line in the file that decides whether any of it ever gets used. A skill that never fires is not a bad skill.

It's an invisible one. Here's why. When you open Hermes, it does not load all of your skills. It cannot. There's not enough room for that. It loads the name and the description of each skill that you have and nothing else. Then it reads those descriptions and decides which one to use. So, the description is not a label. It is the entire trigger. And this is the most common thing wrong with the skills you already have installed.

The description says what the thing is. It never says when to reach for it. So, the agent has nothing to match against. And the skill sits there while you call it by hand and wonder why. The fix is one rule. Say what it does and then say when to use it in the description. So, in this list, we can take Nova as an example, one of my own skills. You are Nova, an agent that handles YouTube content strategy and to end. It says what the skill is, but it doesn't say when to use the skill.

It doesn't help Hermes figure out when to use the skill. So, all we have to do is add in use the skill whenever we're working on a YouTube video. That's it. That's It's literally this simple. Use the skill whenever we're working on a YouTube video. If all of my skills have this sentence when to actually fire, then my agent will just know when it needs to fire any specific skill. Out of the top 55 skills on GitHub that I measured, 49% of them describe what the skill is and never once say when to use it.

And it is the single most common defect in skill building. One more thing here, and it is the opposite of what you would expect. Models under trigger. They skip skills they should have used. So, push the description harder than feels comfortable because the failure you will actually hit is a skill that sits there doing nothing. If I were doing this right now, I would open one skill I installed last week and read its description out loud.

If I cannot tell when it should fire, the agent can't either. And once it fires, the next question is what it costs you every time it does. There is a size where a skill starts costing you more than it gives back. I'm going to open my own file because I'm the worst offender in my own data. This skill right here, Nova, my own YouTube agent, is 19,000 tokens big, for example. The published guidance on this is to keep the body under 500 lines or roughly 5,000 tokens.

So, I am nearly four times over on tokens. And here's the part that actually stings. The median skill in the 55 skills I audited is 1,665 tokens. Mine is 19,000. I'm running something almost 12 times heavier than what everyone else installs. And I built it myself on purpose. Luckily for me, I have my Codex subscription plugged into my Hermes agent, which means it didn't cost me much having this big a skill operate in my workspace daily.

But, this skill sent 6.19 million extra input tokens in that time. And those did not come off my card, they came off my limits. That is the 5-hour window, the 5-hour wall, and that is the part you actually feel if you're also plugged into your own Codex subscription. So, how does a file get like that? It is not typically one bad decision. It is typically 40 small bad decisions. You add a rule because something broke, you add a section because the leading feels risky and adding feels safe.

The structure that stops it is simple. A skill holds two things and only two things. The steps, which is the procedure, and the reference, which is the material the steps need. Once you see it that way, the fix is obvious. Anything the skill only needs sometimes does not belong in the main file. It goes in a reference folder. And the agent opens it only when it actually needs it. That is the single biggest token cut available to you, and almost nobody does it.

One skill out of the top 55 skills on GitHub uses a reference folder. Just 2%. Here is the sentence for that. Read the skill and tell me which parts only get used some of the time. Move those into a references folder and leave a pointer. And there is one part of the body you should never write as instructions at all. Some steps you do not want it thinking about. Here is the distinction, and it is this whole section. If there are lots of right ways to do a step, write instructions and let your agent improvise.

If there's exactly one right way to do something, like you absolutely can't have the possibility of it hallucinating or improvising or doing it any other way, then write code. Because a model does not run your instructions, it reads them and it improvises through them every single time. For a loose step, that is fine, but for arithmetic, it is a coin flip. I mean, look, I asked Hermes three different times to count the number of tokens with instructions only, and I got three completely different answers.

This is why you have to, with anything you can't afford your agent to improvise, to write it as a script that it has to follow step by step. The script never gets loaded into your context, so this is cheaper and more reliable at the same time. And that combination is rare enough that when you find it, you take it. One detail that will bite you, be explicit about whether the agent should run the file or read it. If you leave that to a guess, it will sometimes paste your script into the conversation as reference material and you have just paid tokens for the exact thing you were trying to avoid.

There is also a switch here worth knowing about and it is a safety one. You can mark a skill so the model cannot invoke it and only you can. If a skill sends a message, deploys something, or spends money, that switch is not optional. My rule is that anything with consequences outside my machine waits for me. And now, the file works, which means it is time to make it smaller. The version that works is the one you cut, not the one that you wrote.

There is one test for this and it takes 10 seconds. Delete a paragraph, run the skill again, and if the output did not change, then that paragraph was just doing nothing and you've been paying for it on every single run. I have done this to my own file and it is humbling how much nonsense we can put into it or our own agents can put into it. I mean, whole sections that read like careful instruction and just change nothing because the model already knew that.

Telling it to write a clear summary does not make the summary clearer. It just costs you. Three things that you can hunt for and that's my pro tip for you is the same rule written in two places, so you fix one and the other stays broken. Material that piled up over months because adding felt safer than deleting. And text that sounds like instruction but changes absolutely no behavior whatsoever. Then, the last move is the one that makes a skill worth having.

After every run where the output is not quite right, one question, is this a one-time fix or should this be in the skill forever? If you think it should be in the skill forever, it goes in the file. That is how the thing gets better instead of staying still. My hard rules exist because something broke and I wrote the correction down instead of just fixing it at once. I mean here you could type into your agent, look at the back and forth we just had, what should I add to the skill so we never have this conversation again.

I will be honest about the limit here. This only compounds if you actually do it. And most weeks I forget the skills that got good or the ones I remember to update. It is not clever, it is just repeated. And that is the five, which leaves the question of everything already sitting on your machine. There's one more thing in those 55 skills that I audited and it changed how I install anything online. A skill can ship code that runs.

So, installing a skill from a stranger means putting a stranger's executable next to your files and whatever keys you have lying around. Here's a recent example of a skill I found online that could have gone bad very fast. While I was researching this video, I went to install a well-known agent memory tool that had over 27,000 stars on GitHub. It was MIT licensed, completely legitimate project, nothing wrong with it at all.

The obvious install command, however, pulled a different package entirely because somebody squatted the name on the public registry. The project itself is fine, the command to installing it almost everybody would type was not fine. So, here is what I do now and it takes about 2 minutes. I get the install command from the repository itself, never from a blog post, never from a video, never from, you know, just memory.

Then I send that link to my Hermes agent and ask it to scan for anything that might be malicious about the skill. Never skip this step because, I mean, if someone is squatting that link, or even if that own repo is malicious or has something malicious inside it, you'd rather be safe than sorry. A skill is not a prompt in a folder. It is four decisions. When it fires, which is the description, what goes in it, which is steps and references and nothing else, what gets taken out of it, which is the deletion test, and what happens to the fragile parts, which is a script instead of a guess.

But the reason I actually care about any of this is a lot simpler than that. That job that we just built into a file is a job I'm never going to have to explain again, not next week, not next month, not ever. I do not need to brief it, I do not need to remind it about how to operate this and how to do the skill audit. I do not need to paste the same context in for the fifth time. I just need to call that skill out by name and it just runs.

And that is what people mean when they say an agent feels like an employee. It is not the model getting smarter, it is that the things you used to repeat are now written down once properly and in a place it reads on its own. And once you have five of those, your agent is doing real work for you for every week that you work with it. That is the entire trick and it costs you nothing but the time we just spent building them.

The reason your agent feels inconsistent is usually not the model. It is that you gave it a pile of files nobody ever wrote properly, including myself. The audit skill from this video is free and the file is in the pinned comments. I will upload it on GitHub so you can use it yourself. Download it, point it at your own skills folder, and it will hand you the same table you just watched me build. Comment skill if you enjoyed this video, and make sure to leave a like.

Oh, and if you're new to my channel, then subscribe because I have a ton more content like this coming your way. And, would you look at that? The algorithm gods told me that you're very likely to enjoy this video as well, so click it, and I'll maybe see you there.