Hermes Agent /goal Guide: Use it like THIS!

追加: 2026-05-11

[00:00:00]All right, folks. The last Hermes agent update gave us the slashgoal feature. We briefly covered this in the last update video, but in this video, we're going to dive a bit deeper on how to properly use slashgoal. It's more than just starting off your prompt with slashgoal, build me a million-doll app. That's not really how it works. There's a bunch of steps, bunch of prerequisites that you need to be aware of to make use of the /goal feature. And for those of you who don't know, slashgoal is basically gives your Hermes agent a standing objective that survives across turns. And after every turn, there's this judge that checks whether the goal is satisfied. And if it's not, it will continue feeding a uh continuation prompt back into the same session. So, it's basically the same uh prompt caching as a normal turn. So, it's the same cost uh but with the benefit of your agent being fixated on achieving that goal without any distractions because the biggest problem uh if of not using /goal and how you normally interact with your Hermes agent is if you give it a complex task especially if it has a lot of tool calling involved a lot of API endpoints involved there would be errors along the way that they would uh debug but they would dive down that rabbit hole you know the more errors they have they would dive down that rabbit hole and they would basically forget what their original goal was. Right? So slashgoal basically helps your agent remember every turn what its objective is. And here's an important observation I've made is the slashgoal mostly is used for code refactoring, feature ports, uh test coverage and all you know all sorts of technical stuff.

[00:01:50]And I thought to myself, you know, is this even appropriate for creative projects, right? Is this appropriate for building something that is fun? And the answer to that is yes. Okay. So, the reason why you see /goal skewing towards the technical side uh in their official documentation is because this mechanism is mostly battle tested for the technical stuff, but it is the same concept. As long as you have a clear objective and you understand what are the environments, what are the context you need to provide for your agent, you can still use the slashgoal to build anything you want. Right? So now we're going to show you live on how to build something using the slash goal. Let's see if it actually works. So for this example, uh we're going to build a fighting game. Okay, [laughter] so we're going to build a fighting game.

[00:02:43]And this is the first step that you need to do. is you need to create a project folder or a project directory and this is also one of the benefits of using VPS is in the workspace directory of my Hermes agent uh is it's called claude because I migrated Loki who was my Hermes agent from openclaw back in the early days but I think for your workspace directory the name could be different maybe I think it's Hermes uh but here what you can do is you can just create a folder Uh, and then we'll name it fighting game.

[00:03:19]So, I created the folder inside it's empty. And here is where you want everything related to the fighting game.

[00:03:27]Whether that be files, whether that be materials, sources, all contained here within this folder. Because otherwise, if you type /goal immediately at the start of your prompt in a new chat, your files will be everywhere. And that's a big problem in the future because when you enter a new session with your agent, they don't know which files to reference from. So there's nothing fancy here.

[00:03:49]It's the same philosophy as vibe coding on an IDE. It's just very important to set up an environment first, right? So once we have that folder set up, we can uh CD into that folder. basically just go to that folder and then here we can type Hermes to open up chat with our Hermes agent. Okay, so that's uh step one and then from there I just start chatting with Hermes agent. I describe what my idea is. Uh, and if Hermes has any questions, he can ask me. And once we have a definitive plan, that's when we create the agents.m MD to uh, make sure everything is synced. So, we go back to chat. I tell my agent, I want to build a simple browser-based fighting game. Two characters, basic attacks, health bars, keyboard controls, HTML 5 canvas with JavaScript. Keep it simple, but playable. And I'll show you why. For now, let's just keep it simple. Okay.

[00:04:47]So, go ahead and enter. Okay. So, this is actually impressive. It spun up the agents.mmd by itself. I don't have to create it my my own or I don't have to tell it to create an agents.md. Uh, so this part can be probabilistic because I have, you know, worked with my Hermes agent for about two months now. It understands that every new project I have, uh, it spins up an agents.md. So, that's just part of the self-improving loop of Hermes agent. But if this is maybe like your first time or you barely use uh your Hermes agent then you need to let them know that yeah this is a very important step for every new project. So, it's still thinking, but we can check uh in the file that there is an agents.mmd. That's great. Let's open it. Okay. So, very simple, but we got the overview, text stack, game spec, and the architecture. Okay. Yeah. So, it didn't plan anything. It just went straight to building. I think that's because I didn't tell it to plan first in the first prompt. That was my mistake. But let's actually take a look at the game. So, here's the game. It looks nothing like Street Fighter, but it is, you know, a working game.

[00:05:59]And, you know, S for attack, W for jump, but the enemy is not fighting. So, that's not good. So, this might actually be a good example of what happens if you don't use slash goal. Okay, so this is what I said. This game is not up to my expectation. I'm looking for more of a Street Fighter style game where there are combinations of attacks and the enemy fights back. Look up uh official documentations in the web on how to properly create one and then plan first.

[00:06:28]Do not build yet. So, let's see how that goes. Oh yeah, you can also just use the slash plan command right off the bat um in the beginning. I just found out we actually have the slash plan. I'm actually curious. It's reading part four of this medium.

[00:06:47]I want to make a fighting game. Okay.

[00:06:52]Okay. Okay. So, I'll expect that the game will look something like this. But first, yeah, let a plan first. Okay. So, that took about 14 minutes. My Hermes agent was able to read five research sources and derive seven uh key architectural decisions here. And this is where you can run slashgoal build the fighting game uh according to the plan.

[00:07:14]And if you can see this goal set 20 turn budget, that's when you know you're actually using the slash goal. As the task progress down, you should be able to see one out of 20 uh continuing toward goal. So why there is a 20 turn budget? Uh for those of you who don't know, the uh default for the budget of the uh turns for auto continuing until the goal is accomplished is 20 turns.

[00:07:44]And for the game that we're building, I'm a bit certain that this will take more than 20 turns just because of uh how complex that this project is.

[00:07:52]Because I didn't give my agent a lot of context, a lot of visual documentations or guidelines on how to create this game. I basically just told my agent to research it yourself, plan it out, then build it. Usually, if that's the case, you can assume that you will need more than 20 turns. Remember, this is not it's not your normal turn, okay? That these are turns only if they get something wrong or if they're like stuck in a loop, then they will have the second turn, the third turn, uh, and so on. And if you want to increase your max turn, you can actually do so in the Hermesconfig.yaml file. Um, you should be able to see as you scroll down the goals block uh after you've updated your Hermes agent. But if you don't see this, you can just tell your Hermes agent to add according to the official documentation here. Uh, just to add this to the Hermes config YAML uh, and so on. And under goals, you should see max turns here. The default is set to 20. Uh, and in fact, you can actually raise this. I don't know what the limits are. Certainly, you can't do like I'm sure you cannot do like more than 100, but try maybe raising by an increment of five or 10. So, maybe you can do 25 or maybe you can do 30. That is if you suspect your Hermes agent is dumb and they always get something wrong. But, um, in this case, I'm actually quite surprised. It only took one turn to be able to finish the game here. It made V2. The features have been implemented. We got the combat mechanics. We got blocking. We got the hitbox setup. There's collision layers, too. Uh there's a combo system. That's interesting. Special moves. Uh and most importantly, AI opponent. Damn, there's three modes as well. PVP versus AI easy, AI hard. Okay, this is what I need. Now, let's actually see if it can if it actually uh fits my expectation.

[00:09:49]So yeah, graphics not there, but let's see if the combat is there. So see AD move, S, crouch, W, jump, J, light attack, K heavy attack. Okay, so spec Oh.

[00:10:05][laughter] All right.

[00:10:09]I mean, there's no special graphics or anything. I'm I'm uh But the AI is he even fighting me back? Yeah, there he is. Okay. So, we have a working game.

[00:10:22]It's uh it's accomplished it goal, but it's not, you know, it's not nice.

[00:10:28][laughter] It's not a nice game, but it works. All right. So, this is where um where it's it's good. We can take this iteration and then make a V3. So I think from here on out uh this is where if you don't have the output you're looking for you can always go back to the project directory. All right. And then you can add the resources here your documentations guidelines any files images especially if you have a model that supports image analysis. So Kimmy 2.6 is amazing for front end right. uh here you can add them here. Tell them to reference that uh plan it out and then use the slash uh goal to build it out.

[00:11:17]Right? So I will not use slashgoal here unless we start building the game v3. So everything here that's how I've been using slash goal. Now, I don't know if this is hallucination. Earlier, my agent told me that if the goal is bigger than 20 turns, you can actually use slashgoal sub. But I realize I think my Hermes agent was hallucinating there because there is no slashgo sub. Okay, according to the official documentation.

[00:11:49]So, yeah, just pointing that out there.

[00:11:51]But maybe, hey, what if my agent is anticipating that there might be a slashgo sub? So yeah, I was asking it in the middle when it was building. I was like a bit confused. I was like, there's no slashgoal sub, right? But yeah, that's when I like, okay, you know, whatever. Let's just go with the parent goal. Uh, and it said, you know, I cannot execute the the goal because it was kind of confused with the sub goals.

[00:12:16]Uh, I think it was confusing itself in its memory of this project with my other canban project. So the cananban has that parent child uh link dependency, the parent child task link dependency. So maybe it was like confusing when it was retrieving something from its memory in it. But all in all, uh this is the exact way that you want to use the slashgoal.

[00:12:41]Okay, so bringing it all together, step one is to set up your project. Okay, make a folder in your workspace directory. Then you want to plan first, right? Brainstorm with your Hermes agent. let them ask questions. You can use the slash plan or you can just say in prompt plan this out first. Don't build yet. Uh and then from there if they hadn't created an agents.md for you, you can tell them to create it or you can create it yourself. Put it in the same project folder. And then once you have a plan, that's when you can run /go and then build this out. But so far, I don't really find that much of a difference using /goal with the cananban tool. But I have been using /goal in the middle of my debugging sessions w with the cron job uh cananban uh workflow. So if you haven't seen that video already, I told you that there's no native synergy between cron jobs and canban. So I have my own dduplication loop, but because of that dduplication loop in a new session, my agent messes something up and that's when I need to debug. But there's a lot of errors. So that's when I use, you know, mid session when it's like it's making no progress. It's stuck in a loop. It's asking me clarifying questions and again and again the same clarifying questions. That's when I say, okay, I've had it with this slashgoal fix this error first and then once that error is fixed, okay, the next error fix slashgoal fix error too. So that's how I've been using slashgoal. Uh but yeah, this is one way you can use slashgoal without uh using the cananban feature.

[00:14:17]Uh and I hope that this has helped you out. And just before we go, when not to use slashgoal is for single turn tasks.

[00:14:23]So some uh simple questions, you don't need to use slashgoal. Uh brainstorming.

[00:14:28]Yep. Just like what we said, when you're planning, you don't need to use /goal.

[00:14:31]Uh if you need to um approve iteration manually, you don't use slashgoal because it, you know, they they're in that loop. And then when you're deploying to prod, you don't use slashgoal either. So there's certainly a lot of potentials here for slashgoal.

[00:14:44]I'm actually anticipating there might be a slashgoal subcommand. That would be cool, right? So if there's a very large task, if your agent can determine it's more than 20 turns, they'll you just ask them, well, how many turns does it need for the project to run? So I was indulging its hallucination, and my agent told me it would need five sub goals. It was so convincing too because it's like for the core engine that would take 15 turns. Uh combat that would take 15 turns. Uh special attacks combo system that would take 12 turns. AI opponent would take 10 turns. The UI would take 12 turns. So that ah man if the Hermes agent team watches this uh this would be a very cool implementation. So maybe there's nothing, you know, not all AI eluc hall hallucinations are bad. That can actually spark off some really great ideas. So I hope that this video has helped you on how to properly use slashgo. If you find this video helpful, smash up that like button, subscribe to the channel. My name is Ron. Shannon now.