YouTube2Text

Hermes Agent Skills That Make It 10x More Powerful — Transcript

by AI LABS · 2,587 words · 369 segments · language en · Watch on YouTube

Full transcript

  1. 0:00Ever since Hermes' Agent was released,
  2. 0:02people have been setting up all kinds of
  3. 0:03workflows on it. And to make an already
  4. 0:06powerful Agent even better, people have
  5. 0:08also built a lot of skills for it. Some
  6. 0:10of them have even become some of the
  7. 0:12most trending projects on GitHub. Now,
  8. 0:14you might wonder why you'd need to
  9. 0:15install more skills in the first place
  10. 0:17because Hermes already comes with a lot
  11. 0:19of them, and it can even create new ones
  12. 0:21from your workflows. But, the skills
  13. 0:22Hermes creates only come from the
  14. 0:24processes it learns from you. These
  15. 0:26other skills are built by people who
  16. 0:28have already solved problems Hermes
  17. 0:30can't handle on its own. And they've
  18. 0:31packaged those solutions into skills you
  19. 0:33can use. Now, if you're new to the
  20. 0:34channel, then welcome. We're a software
  21. 0:36company, and this is AI Labs. And in
  22. 0:38this video, we're going to go over the
  23. 0:40skills that actually make using Hermes'
  24. 0:42Agent so much better. But, before we
  25. 0:44start with the first one, you need to
  26. 0:45know this. Even though we're using these
  27. 0:47skills with Hermes, most of them can be
  28. 0:49used with any AI coding agent. Now, if
  29. 0:51you've been working with coding agents,
  30. 0:53even if it's Hermes or literally any
  31. 0:55other agent, they all have the same
  32. 0:57problem. If you're having a long
  33. 0:58conversation, your agent can forget the
  34. 1:00details you mentioned earlier. So, you
  35. 1:02have to give it the same instruction a
  36. 1:03second time. And this happens more often
  37. 1:06when you're using Opus models with
  38. 1:07Hermes because they tend to forget
  39. 1:09things more quickly. This happens
  40. 1:11because there are too many messages in
  41. 1:12your context window, and the model has
  42. 1:14to pay attention to all of them. When
  43. 1:16that happens, it can lose track of the
  44. 1:18smaller details, and you get worse
  45. 1:20answers. To solve that problem, there is
  46. 1:22a skill called Planning with Files. It
  47. 1:24keeps the plan where the agent can't
  48. 1:26ignore it. Instead of keeping the plan
  49. 1:28in the chat and taking up space in the
  50. 1:30context window, this skill converts the
  51. 1:32plan into three files in your project
  52. 1:34folder. The first one is task plan.md,
  53. 1:37which basically splits your main task
  54. 1:39into smaller tasks so the agent can do
  55. 1:41them one at a time. Then, there's
  56. 1:42findings.md, which holds every problem
  57. 1:45the agent faced and how it solved it. If
  58. 1:47the same problem happens again, the
  59. 1:49agent can use those learnings to avoid
  60. 1:51it. And the last one is progress.md,
  61. 1:53which records how far the agent has
  62. 1:55gotten. Now, you might think keeping
  63. 1:57your plan in files isn't new, and most
  64. 1:59of you might already be doing it. What
  65. 2:01sets this skill apart is that the agent
  66. 2:03can't ignore the plan. That's because it
  67. 2:05uses a hook, which forcefully inserts
  68. 2:07the three files into the context window,
  69. 2:09which is why the model can't choose to
  70. 2:11ignore it. Now, most agents don't
  71. 2:13support hooks, but Hermes is one of the
  72. 2:15few that does. And according to their
  73. 2:17own testing, the agent performed much
  74. 2:19better with the hook running because it
  75. 2:20didn't drift away from the task that it
  76. 2:22was assigned. The install commands for
  77. 2:24all the agents are listed in the repo,
  78. 2:26which will be listed in the description
  79. 2:28below. Once you have it installed, it'll
  80. 2:30be added alongside all the other skills.
  81. 2:32And the hooks are going to be
  82. 2:33automatically installed in your agent.
  83. 2:35The skill file explains what the three
  84. 2:37files are, where they should be, and
  85. 2:39what each one does. It also includes all
  86. 2:41the rules the agent needs to follow, so
  87. 2:43it knows what it's working with. So,
  88. 2:45once you have this installed, you can
  89. 2:46run the skill and give it a prompt for
  90. 2:48what you want the agent to do. Here you
  91. 2:50can see that the agent started by
  92. 2:52creating the task plan, the findings,
  93. 2:54and then the progress files as well.
  94. 2:56Then it asked us questions, and we
  95. 2:57answered them as we went. After that
  96. 2:59long session, it finalized the plan and
  97. 3:01gave us the three files. But before we
  98. 3:03go into the next skill, it would be
  99. 3:05great if you subscribe to the channel
  100. 3:07and hit the hype button. This small
  101. 3:09gesture of support goes a long way for
  102. 3:11us. You might already know that you can
  103. 3:13use multiple coding tools with Hermes.
  104. 3:15For example, you can connect your code
  105. 3:17subscription to Hermes agent and use its
  106. 3:19models. You might also know that Hermes
  107. 3:21already has a Claude code skill that
  108. 3:23sends tasks directly to it. And it has
  109. 3:25the same kind of skill for Codex as
  110. 3:26well. But the problem with these skills
  111. 3:28is that Hermes won't use them unless you
  112. 3:30ask it to. You also have to say which
  113. 3:32tasks should go to each agent in your
  114. 3:34prompt. But there is another skill that
  115. 3:36sends a task to any agent that is
  116. 3:38already installed on your system. And
  117. 3:40this skill acts as an orchestrator
  118. 3:42between all of them. It also figures out
  119. 3:44which of your installed coding tools is
  120. 3:46best for which task and gives the work
  121. 3:48to that tool. This is actually a set of
  122. 3:50skills. The main one is called delegate
  123. 3:52setup, which coordinates the other
  124. 3:54agents. It also comes with dedicated
  125. 3:56delegation skills for different coding
  126. 3:57agents, including Claude Code, Codex,
  127. 4:00and Cursor. These agents run in a mode
  128. 4:02where they don't ask permission before
  129. 4:04running any command. So, the coding
  130. 4:05agent doesn't stop and wait for you to
  131. 4:07approve every change. To install this
  132. 4:09skill, run the install command. It asks
  133. 4:11you to choose from 18 skills, so you can
  134. 4:14select the tools installed on your
  135. 4:15system. Once installed, you will see
  136. 4:17that Hermes has the delegate setup skill
  137. 4:19ready for use. To use it, you just have
  138. 4:21to run the skill with the slash command,
  139. 4:23type in the prompt, and give it to
  140. 4:25Hermes agent. First, it identifies which
  141. 4:27coding tools are installed. Then, it
  142. 4:29gives you a complete map of what area
  143. 4:31each agent will handle, along with the
  144. 4:33exact model name and effort level. Once
  145. 4:35you approve the plan, it starts running
  146. 4:37the skills. It then sends each task to
  147. 4:39the agent picked for it. After every
  148. 4:41agent reports its findings, Hermes gives
  149. 4:43you one collective report and marks the
  150. 4:45task as complete. Now, you might already
  151. 4:47know that your Hermes agent runs
  152. 4:49terminal commands. Each terminal command
  153. 4:51is run, it returns an output that the
  154. 4:53agent has to read. 90% of the time, you
  155. 4:56don't even need all of the output that
  156. 4:58command produces. You only need a
  157. 4:59certain part of it. The rest just bloats
  158. 5:01the context for no reason. And just like
  159. 5:03we mentioned before, having too much
  160. 5:05information in the context window makes
  161. 5:07it harder for the agent to focus. That's
  162. 5:09where RTK comes in. It cuts the output
  163. 5:12of the terminal commands the agent runs
  164. 5:14and keeps only the parts the agent
  165. 5:15needs. So, the output gets filtered
  166. 5:17before it ever reaches the context
  167. 5:19window. It has a database of the
  168. 5:21terminal commands agents use most often,
  169. 5:23so it knows how to filter their output.
  170. 5:25For example, if you run tests, the agent
  171. 5:28only needs to see the failed test
  172. 5:29results and not the passing ones. RTK
  173. 5:32does that by default and keeps only the
  174. 5:34failures. But the thing is, it's not
  175. 5:36actually a skill. It's a separate tool
  176. 5:38you install on your device and run from
  177. 5:40the terminal. And it's on this list
  178. 5:42because it makes sure your agent gets
  179. 5:43only the right context. If you want, you
  180. 5:46can also install it directly and tell
  181. 5:48the agent to use it. But, since agents
  182. 5:50tend to forget what we ask, we actually
  183. 5:52created a skill that tells the agent to
  184. 5:54route commands through this tool instead
  185. 5:56of running them the usual way. This
  186. 5:58skill also contains the details the
  187. 5:59agent needs along with a reference for
  188. 6:01each command, so it knows which commands
  189. 6:03it can use and how to use them. You can
  190. 6:05add this skill to your Hermes setup.
  191. 6:07Once you've done that, you can give
  192. 6:09Hermes any task and it will run every
  193. 6:11command through RTK instead of using the
  194. 6:13usual method. Only the output the agent
  195. 6:16needs reaches the context window, so
  196. 6:18Hermes can give you more accurate
  197. 6:19answers. But, before we move on to the
  198. 6:21next one, let's have a word by our
  199. 6:23sponsor, Plural. Running production
  200. 6:25infrastructure usually comes down to one
  201. 6:27thing, the 3:00 a.m. alert and digging
  202. 6:29through Kubernetes logs to find what
  203. 6:31broke. Plural fixes that with
  204. 6:33workbenches. They're AI agents that
  205. 6:35actually operate your infrastructure for
  206. 6:37you and they run inside your existing
  207. 6:38access controls, so nothing happens that
  208. 6:41you haven't already allowed. Here's what
  209. 6:42sold us. We wired it into our
  210. 6:44observability and when an alert fired,
  211. 6:46the agent didn't just ping us. It
  212. 6:48investigated on its own, dug through the
  213. 6:50logs and metrics, and came back with a
  214. 6:52real diagnosis and clear next steps. And
  215. 6:54workbench hooks into the tools you
  216. 6:56already run, like Datadog and Slack, so
  217. 6:58it can kick off on its own the moment an
  218. 7:00alert fires or on a schedule you set.
  219. 7:02And instead of writing prompts from
  220. 7:04scratch, you hand them ready-made skills
  221. 7:06from Plural's library. It's basically an
  222. 7:08on-call engineer that never sleeps. Try
  223. 7:10Plural at the link in the description.
  224. 7:12Now, the [snorts] next skill is by
  225. 7:13Google and it's called Mantis. Mantis is
  226. 7:16a set of security review skills that
  227. 7:18your agents can run. It contains a
  228. 7:20variety of skills, each focusing on a
  229. 7:21different area of security review. This
  230. 7:24setup works no matter what app you're
  231. 7:25building, be it a mobile app or a web
  232. 7:27app. You can install it and use it for
  233. 7:29whatever you're working on. The
  234. 7:30installation command is listed in the
  235. 7:32GitHub repo. When you run the install
  236. 7:34command, you will be asked to choose
  237. 7:35between the many specialized skills it
  238. 7:37has. It's better to install all of them
  239. 7:39because they are separate pieces of the
  240. 7:41whole review process. Once you've
  241. 7:43installed it in your setup, you'll see
  242. 7:45all of the installed listed. For
  243. 7:46example, there is a skill called Mantis
  244. 7:48Plan, and it plans out your security
  245. 7:50review process before you actually
  246. 7:52begin. It also contains a skill called
  247. 7:54Mantis Reflect that goes back over all
  248. 7:56the findings and make sure the failures
  249. 7:58are documented properly. Now, you might
  250. 7:59think if Hermes is an always running
  251. 8:01agent and we don't usually code with it,
  252. 8:03why would we even need to add Mantis
  253. 8:06there? But, the reason is that once your
  254. 8:08app is hosted, it needs security reviews
  255. 8:10again and again, and that isn't
  256. 8:12something you want to remember to do
  257. 8:13yourself. And that's what makes Hermes
  258. 8:15the best place to run them. It has cron
  259. 8:17jobs, which are basically tasks that run
  260. 8:20on their own on a schedule. You can set
  261. 8:22one up that runs this review set against
  262. 8:24the app you've hosted. And because
  263. 8:25Hermes is running all the time, it can
  264. 8:27alert you on Slack or any other
  265. 8:29connected platform if it finds a
  266. 8:30problem. We also have a similar setup
  267. 8:32for our own project. Here, you can see
  268. 8:34that we gave Hermes our project in the
  269. 8:36prompt and told it to run the full set
  270. 8:38over the code. It keeps finding issues
  271. 8:40and applying fixes, which keeps the app
  272. 8:42healthy for us and ensures that it
  273. 8:43doesn't have any bugs. If you've ever
  274. 8:46tried researching with agents, you might
  275. 8:47have run into a huge problem. Your agent
  276. 8:50is denied access to certain content.
  277. 8:52This happens because a lot of sites like
  278. 8:54Reddit don't want models using their
  279. 8:56content. But, those sites are really
  280. 8:57valuable since they have real users
  281. 8:59talking about problems and their
  282. 9:01opinions and experiences, which makes
  283. 9:03the information useful for research. To
  284. 9:05give agents access to that content,
  285. 9:07there is a skill called Agent Reach. It
  286. 9:09gives Hermes access to the sites that
  287. 9:11are usually restricted, including
  288. 9:13Reddit, GitHub, and many others. It uses
  289. 9:15free methods, so it doesn't cost you
  290. 9:17anything. It was also the number one
  291. 9:19trending project on GitHub recently.
  292. 9:21Now, you can install this in your Hermes
  293. 9:23setup, and it'll give you access to more
  294. 9:25than 15 platforms. It can search
  295. 9:27Instagram, Facebook, Twitter, and
  296. 9:29Reddit. Once you have it installed, you
  297. 9:31can use it directly from Hermes and ask
  298. 9:33it to search for any topic on any
  299. 9:35platform. It runs its own scripts, finds
  300. 9:37the relevant information, and then gives
  301. 9:39you the answers that you need. Now, as
  302. 9:41[snorts] you already know, Hermes comes
  303. 9:43with more than 90 skills built in. And
  304. 9:45aside from that, when Hermes sees a
  305. 9:47workflow in your chat that could be
  306. 9:49useful in the future, it actually
  307. 9:50creates a skill out of it, and then it
  308. 9:52can use that skill in future runs. It
  309. 9:54sounds good because it's adapting to you
  310. 9:56automatically, but there is a problem
  311. 9:58with this. The only part of a skill that
  312. 10:00actually goes into the context window is
  313. 10:02its name and description. They get sent
  314. 10:04with each of your messages so that the
  315. 10:06agent knows these skills exist. Now, you
  316. 10:08might say that the name and description
  317. 10:10make up a small chunk and wouldn't
  318. 10:12affect the context window at all. But if
  319. 10:14you have over 100 skills and you've been
  320. 10:16using Hermes for a long time, all those
  321. 10:19names and descriptions start consuming a
  322. 10:21huge portion of the context window. So,
  323. 10:23to solve this issue, there's a whole set
  324. 10:25of skills built to make the agent more
  325. 10:27productive. But the one we're most
  326. 10:29interested in is the skill retriever
  327. 10:31skill. It replaces the full list of
  328. 10:33skill prompts with only the most
  329. 10:35relevant skill descriptions, which
  330. 10:36causes the context size to drop. So,
  331. 10:38basically, it runs another step before
  332. 10:41Hermes calls the model. It splits your
  333. 10:42message into small chunks that a model
  334. 10:45can understand, does the same to every
  335. 10:46skill you have, and then matches your
  336. 10:48message to the skills list. The skills
  337. 10:50that score highest are the ones it
  338. 10:52sends. Their docs put numbers on it.
  339. 10:54They say that with the whole list going
  340. 10:56in, it used around 11,000 tokens. After
  341. 10:58this, it uses only 2.3 thousand, so they
  342. 11:01claim it saves more than 9,000 tokens
  343. 11:03every turn. Now, this helps you in two
  344. 11:05ways. First, there's a lot less going
  345. 11:07into the context window with every
  346. 11:08message you send, so it costs you less
  347. 11:10to run. And second, the agent isn't
  348. 11:12distracted by skills that have nothing
  349. 11:14to do with what you asked, so you get
  350. 11:16better answers. This comes as a plugin,
  351. 11:18and skill retriever is one of the skills
  352. 11:20inside it. Once you've installed the
  353. 11:22plugin, you can enable skill retriever
  354. 11:24from the plugins area. Then the skill
  355. 11:26list going into the context will be much
  356. 11:28shorter than it used to be. Now, the RTK
  357. 11:30skill that we created in this video can
  358. 11:32be found in AI Labs Pro, which is our
  359. 11:34community. So, if you found value in
  360. 11:36what we do and want to support the
  361. 11:37channel, this is the best way to do it.
  362. 11:39The link's in the description. That
  363. 11:41brings us to the end of this video. If
  364. 11:43you'd like to support the channel and
  365. 11:44help us keep making videos like this,
  366. 11:46you can do so by using the Super Thanks
  367. 11:48button below. As always, thank you for
  368. 11:50watching and I'll see you in the next
  369. 11:52one.

About this transcript

This page contains the full transcript of Hermes Agent Skills That Make It 10x More Powerful by AI LABS, generated from the public captions YouTube serves with the video. The transcript has 2,587 words across 369 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.