# How to Build AI Agent Skills That Actually Work (FREE)

Channel: Parker Prompts
Video: https://www.youtube.com/watch?v=roHLWMmDBsc
Duration: 12 min
Language: English
Words: 3225
Transcript page: https://viewrankai.com/tools/youtube-transcript/roHLWMmDBsc

---

[0:00] Building an AI agent that actually works takes exactly three things. A good structure, a server to keep it on, and skills that automate the job. But writing those files by hand is where it falls apart because of how long they take and how precise they have to be. So, I found a free tool that writes all three for you. And by the end of this video, you'll have skills running on your own agent, doing a real job on their own for free. Without skills, an agent works the whole job out from

[0:21] scratch every single time you ask it. You can see it if you run the same request twice. The first answer comes back with a source next to every line and a decision matrix at the bottom. The second one comes back in the same shape, but shorter with no citations anywhere and with different facts. The agent just worked out an approach the first time and that approach died when the run ended. Second round, it worked out a new one. There was no reference to go back to, so there was never anything making the second match the first. And a bigger model doesn't fix that because the model

[0:46] was never the part being inconsistent. It's fine while you're asking one-off questions because you're sat there to read the answer and ask again when it's wrong. It falls apart the moment you want the same job done every week without you watching. And a skill is something that fixes that. It's a reference point for the agent so it can open and follow it instead of working it out again. Physically, it's a folder sat on the agent's machine, and there's one for every skill it has. Inside of it, there's a file called skill.md, which opens with a short description of when this skill applies and holds the steps

[1:14] underneath it. Next to that, you can put script files for the parts of a job that need real code rather than instructions, plus folders for reference material and templates. Though, a simple skill like mine is just the one file. The agent doesn't read any of it until a job matches that description. So, a folder full of skills sits there costing nothing until one of them is wanted, which is what makes them worth building one at a time because they don't get in each other's way and there's no limit you're working towards. The agent I'm doing all of this on is Hermes. It's open- source from an AI lab called Noose

[1:42] Research and this isn't a weekend project. It's sitting at around 236,000 [music] stars on GitHub. And Noose is reportedly raising at a $ 1.5 billion valuation. But here's the part that actually matters for you. Hermes is MIT licensed. Even if news disappeared tomorrow, everything you build stays on your machine and keeps running. So pile up as many skills as you want. Good thing is that it writes skills for itself as well off the back of jobs that took it a lot of steps to get through. Those come out

[2:06] of work, it already got right, though. The jobs you care about are the ones it keeps getting wrong, and those you have to hand it yourself. Every prompt and file I use is in the free doc linked below, so you can follow along on your own job. Before you can hand it anything, we have to do two things first, which acts as the backbone of the entire agent and your skills. The first is the agents identity. And on Hermes, that's one file called soul.md. It's the primary identity for the agent, and it's the first thing that goes into the

[2:29] system prompt on every run it ever does. On this install, it sits at /opt/data/soul.md. And Hermes writes you a starter version the first time it runs. So the file is never missing to begin with. If you leave that starter empty, it falls back to a built-in default, which is why an agent straight out of the box answers like nothing in particular. Whatever is in that file is what the agent believes about itself. It sets the tone, how much it does without checking with you, and what it turns down. And it matters more once skills are involved because the skill tells it how to do the job, while

[2:56] the soul file tells it how to behave the whole time it's doing it. Two agents running the same skill on the same job will hand back different things if their soul files disagree. Writing a page of specific rules about your own working style by hand from an empty file is where this normally stops. So to fix that, I found a free tool called Ordain, and it's accessible at ordain.host, and it writes both the files for you. You put an email in, it sends you a one-time code, and you're through. The site has two builders, one for the soul file and one for skills. And the soul one comes

[3:22] first because a skill needs an agent to work in. Platform goes at the top, so Hermes for me, and openclaw is the other option there. Then there's a name, and that's what it calls itself when you talk to it. Under that is a list of what it handles, and I'm ticking messages and email, research, and summaries, and reminders and scheduling. I'm leaving content creation, customer replies, coding, help, and general assistant switched off because this one isn't doing those. and every extra box you

[3:44] tick makes it less sure what its job is. Personality runs professional, friendly, direct, warm, and playful. I'll select direct. Independence [music] is next, and it's the setting you'll feel most dayto-day. The top one has it ask before it does anything at all. The middle one handles the small stuff on its own and checks in with you on anything big. And the bottom one just gets on with it and tells you after. I'm going with the middle one. Then there's a text box at the bottom, and this is the part I'd spend time on. So, I'm typing that it works in British English and that it gives me the answer first and the

[4:11] reasoning underneath. then that it doesn't pat a list out to hit a round number and when it isn't sure about something it says so be specific in there because vague instructions make a vague agent and this is the file it reads before every job. The mistake to avoid is writing what you'd like it to be rather than how you want it to work since helpful and thorough tell it nothing it doesn't already assume about itself. Then I hit generate and the panel on the right fills in while it writes. What comes out is a page of proper instructions in the agents own voice. So the identity is done and it's

[4:37] sat there on ordained ready to go onto an agent. except there's no agent yet because Hermes needs a machine to run on first. So that's the next thing to sort out. Under the finished soul file, there's a button about needing a server and it goes straight to the hosting or plan I'm running my Hermes on. There are two reasons this isn't optional. The first is that the agent only exists while the machine it's on is switched on. The second one applies once skills are in because a skill is worth most when it runs while you're asleep. The research one I'm about to build runs at 6:00 in the morning and works through a

[5:04] stack of pages on its own and none of that happens on a machine that's shut. Mine runs on Hostinger on the KVM2 plan, which is two processor cores, 8 GB of memory, and 100 GB of storage. That's the one to pick here because Hermes has a one-click install on it. The link in the description drops you straight onto the deploy page. And over on the right side under the coupon box, you'll already see the code Parker applied,

[5:23] taking the final price down for you. KVM2 is already the one selected, so you can leave it exactly as it is. Next is the plan length, and I'd go with 12 months because the discount gets applied on that. Under that, you can turn on daily backups and choose your server's location, and you'll just need to pick the one that's closest to you. Look over the summary on the right, then click

[5:40] continue, pay, and make your account. Once that's done, you'll see the Hermes setup screen, and it hands you a username and a password. You can leave the username as Hermes and copy the password somewhere safe because we're going to need that later. After you hit deploy, Hostinger goes off and builds the server and installs Hermes for you. And a couple of minutes later, your dashboard loads up. On there, Hermes shows up under a part called Docker Manager. You'll also see something called Trafic below it. and you can ignore that as it only handles networking. Your Hermes project will have four random characters on the end of its name, something like Hermes agent

[6:09] ABCD, and you'll need those in just a minute. Up in the top right, there's a button labeled terminal. And clicking it pops open a window called the web terminal. And all we're doing in there is pasting two commands to get Hermes running. The first one opens Hermes itself, and the only part you change is the container name, which is on your dashboard. Paste it, press enter, and now I'm talking to it directly. After that, I'll type Hermes setup, hit enter, and the setup wizard takes over from there. It starts by asking if you want the quick setup or the full one, and full is what I'll choose, so we can get

[6:36] to it with more detail. Its first question is which AI provider you want, and I'm going with Open Router because it has a lot of AI models all under one API key. I'll set that up and put a spending cap of about $25 on it. That cap matters more here than it does in a normal chat because a job running on a schedule can spend money overnight with nobody watching it. Then I connect it to Telegram, which is how I'll talk to it from my phone once I've closed the

[6:56] laptop. Download the soul file as a txt. Then drop it straight into your Telegram chat with Hermes. Don't say anything yet. Just hand it the file. Now jump back to ordain and hit copy [music] prompt. Paste that into Telegram right after the file. That prompt is the instruction set. It tells the agent exactly what to do with the soul file that you just gave it. So I paste that in and Hermes writes its own soul.md into the right folder. Then I start a new session with /reset and it asks me to confirm before it does. So I approve it once. That last step is the one

[7:23] people skip because the soul file is only read when a session starts. So nothing you write applies to the conversation you wrote it in. Ask it who it is now and it answers in the voice I picked with British spelling and no preamble in front of it. It's still working out how to do every job from scratch though and that's the part that's been costing you. Whether a skill works at all comes down to one box in the second builder. So I go back on ordain and into the skill builder platform first and it's the three this time. Openclaw, Hermes, or Claude. It writes to whichever spec you pick

[7:49] because the file format isn't identical across them. Under that is the name which is what the whole folder gets called. So mine is research brief. Then you get to the box marked what should it do. How much you put in here is what decides whether the skill is worth having. Type research topics and give me a summary and you get back a skill that says roughly that. The agent carries on making every decision it was making before only now with a file telling it to. So I'm pasting in the full version instead. That one names the sources it starts from and the order it works through them. It sets what has to be in

[8:16] the finished brief and it says what to do when two sources contradict each other or when it can't find something at all. It's that last instruction that stops it filling the gap with a number that sounds about right. There's a requirements box under that for anything that has to be true about every single run. Mine says every claim carries a link and nothing goes in that it hasn't opened itself. Then I hit generate again and what comes out is a proper skill MD written to the spec of whichever platform I picked. The top of it is the front matter. So the name, a oneline description of when the skill applies

[8:43] and the tags. That description is doing more work than it looks like because it's the only part the agent reads when it's deciding whether to open the rest. Write that description too vaguely and the agent won't open the skill at all, even on the jobs you built it for. Then comes the body, which runs when to use this, the procedure step by step, the pitfalls, and how to check the work at the end. That last part is the one I'd have skipped, and it's why a skill that reads perfectly well still hands back rubbish, because nothing in it ever told the agent what finished looks like. The

[9:08] pitfalls section is doing something similar from the other direction, since it's where you put the mistakes you've already watched it make, and it stops making them once they're written down. You've got two ways to get the finished thing in. Same as before, I upload the txt into Telegram and hit copy prompt. And once that prompt is pasted, the agent writes the folder into hermes/skills on its own or download the zip, put the folder in there directly, and it's live the moment it's in. It reloads the skills itself, and it's picked up from the next session. And because the whole thing is a text file, editing it later means opening it,

[9:35] changing a line, and handing it back to replace the one it's already got. Before trusting it with anything, give it a job you already know the answer to and see whether it followed the procedure or worked around it. It completed the task correctly, but then formatted the answer however it wanted, ignoring the structure from the file entirely. That was the steps being too loose rather than the model being wrong. So, I added a line saying the output goes back in the exact printed format and the second run came back in the right shape. That's how the pitfall section grows. The agent hasn't changed at all. It's just got one

[10:02] job. It no longer has to think its way through, which only counts for anything if what comes back automatically actually holds up. So, it gets the same research job I gave it at the start, word for word. It picks the skill up off the description, opens the procedure, and works on it. The sources it starts from are the ones named in the file, and it goes to the official docs and the change logs before it touches a single forum thread. What comes back opens with a oneline summary, then everything that's changed in the last 90 days with

[10:25] a link sat next to each number. Underneath that are the places where sources disagreed with both positions written out and neither of them dropped. And at the bottom is whatever it couldn't confirm, listed as not found with where it looked, rather than filled in with something that reads about right. I'd have never got that block out of it before because an agent with nothing telling it otherwise would always rather give you an answer than tell you it hasn't got one. Then I run the whole thing again to see what changes. It comes back with the same headings in the same order off the same sources and the only thing that's

[10:52] different between the two runs is what it found. I didn't touch anything in between and that's the whole point of having written the thing down. That second run is the one worth watching because the first one only ever proves it can do the job once which it could do before any of this. The last step is

[11:05] telling it when to do that on its own. So, I ask it to run this every weekday at 6:00 and message me the brief on Telegram and it sets that [music] up itself. Four skills are worth building before anything else and they're the jobs you already do by hand every week. Research is the one I've just done and it pays back fastest because it's the most repetitive thing on the list. The second one is your inbox where it drafts replies in your voice and leaves them sitting for you to send. So, you're

[11:26] editing instead of writing from nothing. After that, there's content where you give it a topic and it comes back with a script in the structure you use every week rather than the structure a model assumes a script has. And the last one is outreach, where it finds people worth contacting and writes the first message, which is the part I never get round to doing myself. For picking your own, anything you've done more than three times the same way is a skill. Anything where you take a different approach depending on what you find in the first 10 minutes isn't, and forcing it into one makes the agent worse because you've

[11:52] handed it steps that don't fit the job in front of it. You can stack as many as you like since it only opens the ones a job matches, so the folder can get big without any of it slowing a rundown. And the more of them you write, the more of your own working style is sat in files instead of in your head, which is the point at which the agent stops being a chat window you visit. So the job that used to come back different every time

[12:10] now comes back the same way every time. The run I just showed you happened automatically while I was asleep. And that can only happen when your agent is actually hosted at a server that never sleeps. So grab the Hostinger or KVM2 plan through the first link in the description with the code Parker and put your first skill to work. Thank you for watching and I'll see you in the next

---

About this transcript
Read from YouTube's own caption track and laid out by ViewRank AI (https://viewrankai.com).
ViewRank AI finds the videos already beating a creator's own average on Instagram, TikTok and YouTube Shorts, transcribes them from the audio itself in more than 60 languages, and turns what worked into new ideas and scripts.
Free transcript tools, no account needed: https://viewrankai.com/tools
How to read any video this way: https://viewrankai.com/llms.txt