YouTube2Text

Anthropic Engineer Explains: What to Build Instead of AI Agents — Transcript

by Nate Herk | AI Automation · 2,523 words · 374 segments · language en · Watch on YouTube

Full transcript

  1. 0:00So Anthropic engineers just said that
  2. 0:01they stopped building agents and they
  3. 0:03started building something completely
  4. 0:04different. So if you're still focused on
  5. 0:05building agents, you're probably wasting
  6. 0:07hours on something that will never work
  7. 0:08the way that you actually want it to.
  8. 0:09But if you focus on what these engineers
  9. 0:11are actually building, you'll have a
  10. 0:12system that improves on its own. So in
  11. 0:14this video, I'm going to show you what
  12. 0:15they said to build instead and the four
  13. 0:17things that actually make it work so
  14. 0:18that you can start getting the same
  15. 0:19quality results that Anthropic engineers
  16. 0:21are getting. So let's get into it. All
  17. 0:22right, so Anthropic isn't saying that
  18. 0:24agents are dead. Barry Zhang and Mahesh
  19. 0:26Maragh, the two people who created agent
  20. 0:28skills at Anthropic, said that they
  21. 0:29basically stopped rebuilding a separate
  22. 0:31agent for every single job because the
  23. 0:33agent underneath had become way more
  24. 0:35general purpose than they expected. So
  25. 0:36real quick, the easiest way to
  26. 0:37understand this is to look at your
  27. 0:39phone. Your phone has a processor, an
  28. 0:41operating system, and then all the apps
  29. 0:43that you actually use every day. And a
  30. 0:44few massive companies build the
  31. 0:46processor and the operating system. You
  32. 0:47probably aren't changing either one of
  33. 0:49those things, but you can choose the
  34. 0:51apps and each app gives that same phone
  35. 0:53a very specific capability. And the
  36. 0:54whole AI stack is starting to look
  37. 0:56pretty similar to that. The model is
  38. 0:58kind of like the processor, the agent
  39. 0:59runtime is like the operating system,
  40. 1:00[music]
  41. 1:01and then skills are the apps. So Claude
  42. 1:03code can already read files, write code,
  43. 1:05call tools, and work through a task. You
  44. 1:07don't need to necessarily rebuild all of
  45. 1:08that every time that you want help
  46. 1:10creating a presentation or researching a
  47. 1:11company or writing a LinkedIn post or
  48. 1:13doing whatever kind of things you need
  49. 1:14to do like that. You give the same
  50. 1:16general purpose agent a skill that
  51. 1:17contains the process, the context, the
  52. 1:20scripts, and the examples for that
  53. 1:21specific job. So now let me talk about
  54. 1:23the four practical ways to make those
  55. 1:25skills work better. So the first one is
  56. 1:26simple. Stop making Claude solve the
  57. 1:28same technical problem over and over.
  58. 1:30Anthropic's team kept watching Claude
  59. 1:32write basically the exact same Python
  60. 1:33script every time it needed to apply
  61. 1:35styling to a slide deck. It would spend
  62. 1:37tons of tokens just recreating the code
  63. 1:39that has already been written and
  64. 1:40because it was rebuilding it from
  65. 1:41scratch, the result could change from
  66. 1:42one run to the next, so it wasn't super
  67. 1:44consistent. So what they did is they had
  68. 1:46Claude save that script inside the skill
  69. 1:48as, in their own words, a tool for its
  70. 1:50future self. So now the next time it
  71. 1:52needs to style a presentation, it can
  72. 1:53just run the version that already was
  73. 1:55proven to work. And developers have
  74. 1:56followed this idea forever. It's called
  75. 1:58DRY or DRY, which stands for don't
  76. 2:00yourself. If you solve a problem in
  77. 2:02code, you can just save the solution and
  78. 2:03reuse it instead of rewriting the code
  79. 2:05every single time. And you can do the
  80. 2:06exact same thing with Claude. I've got
  81. 2:08skills in my own AI operating system
  82. 2:09that use the same renderers, the same
  83. 2:11templates, scripts, and stuff like that
  84. 2:13every time. My carousel workflow doesn't
  85. 2:14ask Claude to reinvent how a slide gets
  86. 2:16rendered on every run. The skill just
  87. 2:17points it to the files that already
  88. 2:19work, and then Claude can just focus on
  89. 2:20the new content like literally just
  90. 2:21changing the text. So, the next time
  91. 2:23Claude writes a script that gives you a
  92. 2:24result that you really like, don't leave
  93. 2:26that code trapped inside the chat just
  94. 2:28to get lost later when you start up new
  95. 2:29sessions. Tell it something like, "Save
  96. 2:31the script you just used inside the
  97. 2:32skills script folder. Update the skill
  98. 2:34that I'm doing so that future runs
  99. 2:35execute that file instead of trying to
  100. 2:37rewrite it again." Then I want you to
  101. 2:38run the skill again and verify the
  102. 2:40results. And real quick, you still
  103. 2:41obviously need to test it. You run the
  104. 2:43same type of task twice, you compare the
  105. 2:44important parts, and make sure the skill
  106. 2:46is actually calling the saved file. The
  107. 2:48surrounding AI output may still vary a
  108. 2:50little bit, but you've replaced one
  109. 2:51fresh guess with a proven piece of code.
  110. 2:53So, moral of the story, don't pay Claude
  111. 2:55to rediscover a solution that you
  112. 2:56already have. But, once you start
  113. 2:58building a bunch of these skills, Claude
  114. 3:00needs to know which one belongs to the
  115. 3:01job in front of it. So, let's move on to
  116. 3:03number two. So, think about a mechanic
  117. 3:04for a second. A mechanic might own
  118. 3:06hundreds of tools, but he doesn't dump
  119. 3:07every single one onto the bench before
  120. 3:09he starts changing a tire. He basically
  121. 3:11just identifies the job and grabs the
  122. 3:13few tools that he needs and leaves
  123. 3:14everything else inside the toolbox. And
  124. 3:16skills work in a very similar way
  125. 3:17because of something that Anthropic
  126. 3:19calls progressive disclosure. So, when
  127. 3:21Claude starts, it doesn't read the full
  128. 3:22instructions and all the examples and
  129. 3:24all the scripts from every single skill
  130. 3:26you have. That would be a huge waste of
  131. 3:27time and tokens. It starts with the name
  132. 3:29and the description of each skill. And
  133. 3:31this is called the YAML front matter.
  134. 3:32Then, when your specific prompt matches
  135. 3:34a description of a skill, that's when it
  136. 3:36reads the full skill.md file. And any
  137. 3:37larger references or scripts can just
  138. 3:39stay in the folder until the task
  139. 3:40actually needs them. Anthropic describes
  140. 3:42this as letting Claude load information
  141. 3:44only as needed. That basically keeps
  142. 3:46irrelevant instructions out of the
  143. 3:48working context, which helps you avoid
  144. 3:49the bloat and confusion, which sometimes
  145. 3:51gets called context rot. But, this
  146. 3:52depends on your description being
  147. 3:53really, really clear. If one skill says
  148. 3:55help with content and another one says
  149. 3:57create marketing assets, then Claude's
  150. 3:58basically guessing because those
  151. 3:59descriptions sort of overlap, and they
  152. 4:01don't tell the agent when either of
  153. 4:02those skills should actually run. So, a
  154. 4:04stronger description might say something
  155. 4:05like, "This skill creates LinkedIn
  156. 4:07carousels from a topic, from a
  157. 4:08transcript, or an outline. Use this when
  158. 4:10the user asks for a carousel, carousel
  159. 4:12slides, or a LinkedIn document post."
  160. 4:14Now, Claude knows what the skill does
  161. 4:15and exactly when to use it. So, just
  162. 4:17keep each skill focused on one specific
  163. 4:19job, put the words a real person would
  164. 4:20use inside the description, and make
  165. 4:22sure two skills aren't competing for the
  166. 4:24same request. And you can actually have
  167. 4:25Claude code audit this for you.
  168. 4:27>> [music]
  169. 4:27>> Just say, "Hey, review all my skill
  170. 4:28descriptions. For each one, tell me what
  171. 4:30it does, when it should trigger, and
  172. 4:32where it overlaps with any other skills.
  173. 4:33Rewrite only the descriptions that are
  174. 4:35[music] ambiguous." And then you can
  175. 4:36test these three things. The first one
  176. 4:38is an obvious request that should
  177. 4:39trigger it. The second one is a
  178. 4:40differently worded request that still
  179. 4:41should trigger it. And the third one is
  180. 4:43an unrelated request that definitely
  181. 4:44should not trigger it. And just remember
  182. 4:45that a skill that Claude can't find is
  183. 4:47basically a skill that you don't have.
  184. 4:48So, finding the correct skill handles
  185. 4:50today's tasks. But the next step is
  186. 4:52actually making sure that the skill gets
  187. 4:53better every single time you use it. So,
  188. 4:55number three, every time you correct
  189. 4:56Claude and then you close the chat,
  190. 4:58there's a pretty good chance you just
  191. 4:59threw that lesson away. Maybe [music] it
  192. 5:00used the wrong tone, or it skipped a
  193. 5:02validation step, or maybe it formatted
  194. 5:04the final output in a way that you don't
  195. 5:06want to see again. And if all you say to
  196. 5:07Claude is, "Fix it," then it probably
  197. 5:08will fix it, but the process stays
  198. 5:10broken. And this is something I've had
  199. 5:11to work through inside my own AI
  200. 5:12operating system. If an agent tells me
  201. 5:14it can't find a file that I know exists,
  202. 5:15I don't just hand it the path and keep
  203. 5:17moving. I ask it to backtrack. I
  204. 5:18basically ask it to show me its work.
  205. 5:20Show me where it searched. Figure out
  206. 5:21why it missed the file. And then update
  207. 5:23the routing or the skill so that next
  208. 5:25run starts in the correct place.
  209. 5:27Anthropic designed skills as a step
  210. 5:28toward this kind of continuous learning.
  211. 5:31Their guarantee is that anything that
  212. 5:32Claude writes down can be used
  213. 5:33efficiently by a future version of
  214. 5:35itself. Now, here's the thing. Skills
  215. 5:36don't remember every single thing, and
  216. 5:38they aren't a recording of every
  217. 5:39conversation that you've ever had. They
  218. 5:40basically store the procedural knowledge
  219. 5:42that Claude needs to do the specific
  220. 5:43[music] job. So, when a process is
  221. 5:45wrong, update the instructions in the
  222. 5:47skill.md. When Claude is missing your
  223. 5:48voice, or your brand, or your examples,
  224. 5:51then add a reference file. And when the
  225. 5:52same mistake keeps happening, then add a
  226. 5:53clear rule that explicitly prevents
  227. 5:55that. And then rerun the same task. You
  228. 5:57can use a prompt like this. Review what
  229. 5:59went wrong during this run. Decide
  230. 6:00whether the cause was the process, or
  231. 6:02missing context, or a weak rule, or
  232. 6:04unreliable code. And then, update the
  233. 6:06skill in the smallest durable place,
  234. 6:08then rerun the same task, and verify
  235. 6:10[music] the fix. Over time, the skill
  236. 6:11becomes a living record of how you want
  237. 6:13the job done. And I do want to be
  238. 6:14precise about the phrase model proof. No
  239. 6:16skill can just force a weaker model to
  240. 6:18perform exactly like the stronger one.
  241. 6:20Different models will obviously still
  242. 6:21produce different results, and they
  243. 6:22interpret skills in different ways. But,
  244. 6:24your process can be portable. Agent
  245. 6:26skills are an open format, so the same
  246. 6:28core like skills folder can work across
  247. 6:30compatible agent harnesses like Codex or
  248. 6:32Hermes Agent or anything else. So, test
  249. 6:34an important skill with another
  250. 6:35compatible agent. If the result falls
  251. 6:36apart, then look for hidden assumptions,
  252. 6:38missing examples, or instructions that
  253. 6:40only one model understands, and then
  254. 6:41tighten up the skill and keep testing.
  255. 6:43All right, so the first three were very
  256. 6:44important, but this one is probably the
  257. 6:46most important. A skill shouldn't hand
  258. 6:48you its first attempt and call the job
  259. 6:49done. This is probably the biggest gap
  260. 6:51that I see with a lot of AI workflows.
  261. 6:53You run the skill, it creates the thing,
  262. 6:54saves the file, and then comes back to
  263. 6:55you and says, "Hey, I'm done." And then
  264. 6:57you open it, and you realize that the
  265. 6:58formatting's broken, the sources don't
  266. 7:00support the claims, or that the script
  267. 7:01falls completely flat for the person
  268. 7:03you're trying to reach. So, the AI did
  269. 7:04maybe 70 or 80% of the job, and now
  270. 7:07you're the human, you know, manually
  271. 7:09doing the last 20 or 30%. But, if you
  272. 7:11already know how to personally check
  273. 7:12that work, then just bake those checks
  274. 7:14into the skill, and let the AI close
  275. 7:15more of that gap for you. So, for
  276. 7:17example, for a slide deck, have the
  277. 7:18skill render each slide as an image,
  278. 7:20inspect the screenshots, fix anything
  279. 7:22that's cropped or hard to read or out of
  280. 7:23bounds, and render it. For a research
  281. 7:25report, make it open the primary
  282. 7:26sources, match the claims [music] to the
  283. 7:28evidence, and remove anything that it
  284. 7:29couldn't actually verify. And if you're
  285. 7:31creating a script or an ad, or something
  286. 7:33more subjective, then have a few
  287. 7:34different personas, like a few different
  288. 7:36sub agents review it and discuss it. A
  289. 7:38beginner agent can tell you where
  290. 7:39they're confused. [music] A skeptical
  291. 7:40buyer agent can tell you what they don't
  292. 7:42believe. Someone from your actual
  293. 7:43audience can tell you where they would
  294. 7:45probably click away. Now, you don't have
  295. 7:46to accept every single piece of feedback
  296. 7:47that you get from these different agent
  297. 7:49personas. That would probably make the
  298. 7:50output worse, but the skill can find the
  299. 7:52issues that show up more than once, make
  300. 7:53the strongest revisions, and then run
  301. 7:55the review again. And sometimes just
  302. 7:56hearing those different perspectives is
  303. 7:58really helpful. And real quick,
  304. 7:59verification isn't Claude reading its
  305. 8:00own work and saying, "Looks good to me."
  306. 8:02It needs some kind of evidence outside
  307. 8:04of that first draft, whether whether a
  308. 8:05screenshot, a test result, [music] a
  309. 8:06source, a reference example, or like I
  310. 8:08said, feedback from a few different
  311. 8:09agent perspectives. And you can add
  312. 8:11something like this into almost any
  313. 8:12skill. Say something like, "Before
  314. 8:14returning the final output, define the
  315. 8:15acceptance criteria, create the first
  316. 8:16version, inspect it using the relevant
  317. 8:18verification method, fix every issue you
  318. 8:20find, and then run another pass. Return
  319. 8:22the output only after it meets the
  320. 8:23criteria with a short summary of what
  321. 8:25you checked. And if something can't be
  322. 8:26verified, then tell me exactly what
  323. 8:28remains." It's even better if there's
  324. 8:29some sort of objective success metric
  325. 8:32that you can set and just have the
  326. 8:33agents keep working until it hits that
  327. 8:35objectively. But anyways, now the first
  328. 8:36output is basically an internal draft.
  329. 8:38Claude reviews it, catches the obvious
  330. 8:40problems, and then improves it before
  331. 8:41you see any of it, before you waste any
  332. 8:43of your time and attention on it.
  333. 8:45Because you're ultimately still going to
  334. 8:46be the final judge, especially when
  335. 8:47there's things like taste or strategy or
  336. 8:49business judgment involved in the
  337. 8:50process. The goal though is just to stop
  338. 8:52spending your time catching problems
  339. 8:54that the AI could have caught on its
  340. 8:55own. Your first look should not be the
  341. 8:57agent's first look. It should be the
  342. 8:58agent's, you know, fourth or fifth or
  343. 9:00maybe even sixth look. So now you know
  344. 9:01what Anthropic engineers are actually
  345. 9:03building. They save proven code instead
  346. 9:05of rewriting it. They give every skill a
  347. 9:06precise description so Claude loads only
  348. 9:08the correct one. They turn corrections
  349. 9:10into durable instructions that keep
  350. 9:11improving, and they verify the work
  351. 9:13before it ever reaches you. That's how
  352. 9:15you take a general purpose agent and you
  353. 9:16teach it how you work specifically. And
  354. 9:18if you want to see how I organize the
  355. 9:19agents, the context, connections,
  356. 9:21capabilities, and cadence around all of
  357. 9:22this, then I'll put my full AI operating
  358. 9:24system course on screen right up here. I
  359. 9:27also have free courses, templates, and
  360. 9:28resources inside my free community that
  361. 9:30will help you build agents and workflows
  362. 9:32from scratch. You can join with the link
  363. 9:33in the description, and if you want to
  364. 9:34go deeper [music] on turning these
  365. 9:35skills into a career, then you can check
  366. 9:37out my plus community. But anyways, that
  367. 9:38is going to do it for this one. So if
  368. 9:40you guys enjoyed it or you learned
  369. 9:40something new, please give it a like, it
  370. 9:42helps me out a ton. And as always, I
  371. 9:43appreciate you guys making it to the end
  372. 9:44of the video. I'll see you all in the
  373. 9:45next one.
  374. 9:46Thanks, everyone.

About this transcript

This page contains the full transcript of Anthropic Engineer Explains: What to Build Instead of AI Agents by Nate Herk | AI Automation, generated from the public captions YouTube serves with the video. The transcript has 2,523 words across 374 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.