Anthropic Engineer Explains: What to Build Instead of AI Agents — Transcript
Full transcript
- 0:00So Anthropic engineers just said that
- 0:01they stopped building agents and they
- 0:03started building something completely
- 0:04different. So if you're still focused on
- 0:05building agents, you're probably wasting
- 0:07hours on something that will never work
- 0:08the way that you actually want it to.
- 0:09But if you focus on what these engineers
- 0:11are actually building, you'll have a
- 0:12system that improves on its own. So in
- 0:14this video, I'm going to show you what
- 0:15they said to build instead and the four
- 0:17things that actually make it work so
- 0:18that you can start getting the same
- 0:19quality results that Anthropic engineers
- 0:21are getting. So let's get into it. All
- 0:22right, so Anthropic isn't saying that
- 0:24agents are dead. Barry Zhang and Mahesh
- 0:26Maragh, the two people who created agent
- 0:28skills at Anthropic, said that they
- 0:29basically stopped rebuilding a separate
- 0:31agent for every single job because the
- 0:33agent underneath had become way more
- 0:35general purpose than they expected. So
- 0:36real quick, the easiest way to
- 0:37understand this is to look at your
- 0:39phone. Your phone has a processor, an
- 0:41operating system, and then all the apps
- 0:43that you actually use every day. And a
- 0:44few massive companies build the
- 0:46processor and the operating system. You
- 0:47probably aren't changing either one of
- 0:49those things, but you can choose the
- 0:51apps and each app gives that same phone
- 0:53a very specific capability. And the
- 0:54whole AI stack is starting to look
- 0:56pretty similar to that. The model is
- 0:58kind of like the processor, the agent
- 0:59runtime is like the operating system,
- 1:00[music]
- 1:01and then skills are the apps. So Claude
- 1:03code can already read files, write code,
- 1:05call tools, and work through a task. You
- 1:07don't need to necessarily rebuild all of
- 1:08that every time that you want help
- 1:10creating a presentation or researching a
- 1:11company or writing a LinkedIn post or
- 1:13doing whatever kind of things you need
- 1:14to do like that. You give the same
- 1:16general purpose agent a skill that
- 1:17contains the process, the context, the
- 1:20scripts, and the examples for that
- 1:21specific job. So now let me talk about
- 1:23the four practical ways to make those
- 1:25skills work better. So the first one is
- 1:26simple. Stop making Claude solve the
- 1:28same technical problem over and over.
- 1:30Anthropic's team kept watching Claude
- 1:32write basically the exact same Python
- 1:33script every time it needed to apply
- 1:35styling to a slide deck. It would spend
- 1:37tons of tokens just recreating the code
- 1:39that has already been written and
- 1:40because it was rebuilding it from
- 1:41scratch, the result could change from
- 1:42one run to the next, so it wasn't super
- 1:44consistent. So what they did is they had
- 1:46Claude save that script inside the skill
- 1:48as, in their own words, a tool for its
- 1:50future self. So now the next time it
- 1:52needs to style a presentation, it can
- 1:53just run the version that already was
- 1:55proven to work. And developers have
- 1:56followed this idea forever. It's called
- 1:58DRY or DRY, which stands for don't
- 2:00yourself. If you solve a problem in
- 2:02code, you can just save the solution and
- 2:03reuse it instead of rewriting the code
- 2:05every single time. And you can do the
- 2:06exact same thing with Claude. I've got
- 2:08skills in my own AI operating system
- 2:09that use the same renderers, the same
- 2:11templates, scripts, and stuff like that
- 2:13every time. My carousel workflow doesn't
- 2:14ask Claude to reinvent how a slide gets
- 2:16rendered on every run. The skill just
- 2:17points it to the files that already
- 2:19work, and then Claude can just focus on
- 2:20the new content like literally just
- 2:21changing the text. So, the next time
- 2:23Claude writes a script that gives you a
- 2:24result that you really like, don't leave
- 2:26that code trapped inside the chat just
- 2:28to get lost later when you start up new
- 2:29sessions. Tell it something like, "Save
- 2:31the script you just used inside the
- 2:32skills script folder. Update the skill
- 2:34that I'm doing so that future runs
- 2:35execute that file instead of trying to
- 2:37rewrite it again." Then I want you to
- 2:38run the skill again and verify the
- 2:40results. And real quick, you still
- 2:41obviously need to test it. You run the
- 2:43same type of task twice, you compare the
- 2:44important parts, and make sure the skill
- 2:46is actually calling the saved file. The
- 2:48surrounding AI output may still vary a
- 2:50little bit, but you've replaced one
- 2:51fresh guess with a proven piece of code.
- 2:53So, moral of the story, don't pay Claude
- 2:55to rediscover a solution that you
- 2:56already have. But, once you start
- 2:58building a bunch of these skills, Claude
- 3:00needs to know which one belongs to the
- 3:01job in front of it. So, let's move on to
- 3:03number two. So, think about a mechanic
- 3:04for a second. A mechanic might own
- 3:06hundreds of tools, but he doesn't dump
- 3:07every single one onto the bench before
- 3:09he starts changing a tire. He basically
- 3:11just identifies the job and grabs the
- 3:13few tools that he needs and leaves
- 3:14everything else inside the toolbox. And
- 3:16skills work in a very similar way
- 3:17because of something that Anthropic
- 3:19calls progressive disclosure. So, when
- 3:21Claude starts, it doesn't read the full
- 3:22instructions and all the examples and
- 3:24all the scripts from every single skill
- 3:26you have. That would be a huge waste of
- 3:27time and tokens. It starts with the name
- 3:29and the description of each skill. And
- 3:31this is called the YAML front matter.
- 3:32Then, when your specific prompt matches
- 3:34a description of a skill, that's when it
- 3:36reads the full skill.md file. And any
- 3:37larger references or scripts can just
- 3:39stay in the folder until the task
- 3:40actually needs them. Anthropic describes
- 3:42this as letting Claude load information
- 3:44only as needed. That basically keeps
- 3:46irrelevant instructions out of the
- 3:48working context, which helps you avoid
- 3:49the bloat and confusion, which sometimes
- 3:51gets called context rot. But, this
- 3:52depends on your description being
- 3:53really, really clear. If one skill says
- 3:55help with content and another one says
- 3:57create marketing assets, then Claude's
- 3:58basically guessing because those
- 3:59descriptions sort of overlap, and they
- 4:01don't tell the agent when either of
- 4:02those skills should actually run. So, a
- 4:04stronger description might say something
- 4:05like, "This skill creates LinkedIn
- 4:07carousels from a topic, from a
- 4:08transcript, or an outline. Use this when
- 4:10the user asks for a carousel, carousel
- 4:12slides, or a LinkedIn document post."
- 4:14Now, Claude knows what the skill does
- 4:15and exactly when to use it. So, just
- 4:17keep each skill focused on one specific
- 4:19job, put the words a real person would
- 4:20use inside the description, and make
- 4:22sure two skills aren't competing for the
- 4:24same request. And you can actually have
- 4:25Claude code audit this for you.
- 4:27>> [music]
- 4:27>> Just say, "Hey, review all my skill
- 4:28descriptions. For each one, tell me what
- 4:30it does, when it should trigger, and
- 4:32where it overlaps with any other skills.
- 4:33Rewrite only the descriptions that are
- 4:35[music] ambiguous." And then you can
- 4:36test these three things. The first one
- 4:38is an obvious request that should
- 4:39trigger it. The second one is a
- 4:40differently worded request that still
- 4:41should trigger it. And the third one is
- 4:43an unrelated request that definitely
- 4:44should not trigger it. And just remember
- 4:45that a skill that Claude can't find is
- 4:47basically a skill that you don't have.
- 4:48So, finding the correct skill handles
- 4:50today's tasks. But the next step is
- 4:52actually making sure that the skill gets
- 4:53better every single time you use it. So,
- 4:55number three, every time you correct
- 4:56Claude and then you close the chat,
- 4:58there's a pretty good chance you just
- 4:59threw that lesson away. Maybe [music] it
- 5:00used the wrong tone, or it skipped a
- 5:02validation step, or maybe it formatted
- 5:04the final output in a way that you don't
- 5:06want to see again. And if all you say to
- 5:07Claude is, "Fix it," then it probably
- 5:08will fix it, but the process stays
- 5:10broken. And this is something I've had
- 5:11to work through inside my own AI
- 5:12operating system. If an agent tells me
- 5:14it can't find a file that I know exists,
- 5:15I don't just hand it the path and keep
- 5:17moving. I ask it to backtrack. I
- 5:18basically ask it to show me its work.
- 5:20Show me where it searched. Figure out
- 5:21why it missed the file. And then update
- 5:23the routing or the skill so that next
- 5:25run starts in the correct place.
- 5:27Anthropic designed skills as a step
- 5:28toward this kind of continuous learning.
- 5:31Their guarantee is that anything that
- 5:32Claude writes down can be used
- 5:33efficiently by a future version of
- 5:35itself. Now, here's the thing. Skills
- 5:36don't remember every single thing, and
- 5:38they aren't a recording of every
- 5:39conversation that you've ever had. They
- 5:40basically store the procedural knowledge
- 5:42that Claude needs to do the specific
- 5:43[music] job. So, when a process is
- 5:45wrong, update the instructions in the
- 5:47skill.md. When Claude is missing your
- 5:48voice, or your brand, or your examples,
- 5:51then add a reference file. And when the
- 5:52same mistake keeps happening, then add a
- 5:53clear rule that explicitly prevents
- 5:55that. And then rerun the same task. You
- 5:57can use a prompt like this. Review what
- 5:59went wrong during this run. Decide
- 6:00whether the cause was the process, or
- 6:02missing context, or a weak rule, or
- 6:04unreliable code. And then, update the
- 6:06skill in the smallest durable place,
- 6:08then rerun the same task, and verify
- 6:10[music] the fix. Over time, the skill
- 6:11becomes a living record of how you want
- 6:13the job done. And I do want to be
- 6:14precise about the phrase model proof. No
- 6:16skill can just force a weaker model to
- 6:18perform exactly like the stronger one.
- 6:20Different models will obviously still
- 6:21produce different results, and they
- 6:22interpret skills in different ways. But,
- 6:24your process can be portable. Agent
- 6:26skills are an open format, so the same
- 6:28core like skills folder can work across
- 6:30compatible agent harnesses like Codex or
- 6:32Hermes Agent or anything else. So, test
- 6:34an important skill with another
- 6:35compatible agent. If the result falls
- 6:36apart, then look for hidden assumptions,
- 6:38missing examples, or instructions that
- 6:40only one model understands, and then
- 6:41tighten up the skill and keep testing.
- 6:43All right, so the first three were very
- 6:44important, but this one is probably the
- 6:46most important. A skill shouldn't hand
- 6:48you its first attempt and call the job
- 6:49done. This is probably the biggest gap
- 6:51that I see with a lot of AI workflows.
- 6:53You run the skill, it creates the thing,
- 6:54saves the file, and then comes back to
- 6:55you and says, "Hey, I'm done." And then
- 6:57you open it, and you realize that the
- 6:58formatting's broken, the sources don't
- 7:00support the claims, or that the script
- 7:01falls completely flat for the person
- 7:03you're trying to reach. So, the AI did
- 7:04maybe 70 or 80% of the job, and now
- 7:07you're the human, you know, manually
- 7:09doing the last 20 or 30%. But, if you
- 7:11already know how to personally check
- 7:12that work, then just bake those checks
- 7:14into the skill, and let the AI close
- 7:15more of that gap for you. So, for
- 7:17example, for a slide deck, have the
- 7:18skill render each slide as an image,
- 7:20inspect the screenshots, fix anything
- 7:22that's cropped or hard to read or out of
- 7:23bounds, and render it. For a research
- 7:25report, make it open the primary
- 7:26sources, match the claims [music] to the
- 7:28evidence, and remove anything that it
- 7:29couldn't actually verify. And if you're
- 7:31creating a script or an ad, or something
- 7:33more subjective, then have a few
- 7:34different personas, like a few different
- 7:36sub agents review it and discuss it. A
- 7:38beginner agent can tell you where
- 7:39they're confused. [music] A skeptical
- 7:40buyer agent can tell you what they don't
- 7:42believe. Someone from your actual
- 7:43audience can tell you where they would
- 7:45probably click away. Now, you don't have
- 7:46to accept every single piece of feedback
- 7:47that you get from these different agent
- 7:49personas. That would probably make the
- 7:50output worse, but the skill can find the
- 7:52issues that show up more than once, make
- 7:53the strongest revisions, and then run
- 7:55the review again. And sometimes just
- 7:56hearing those different perspectives is
- 7:58really helpful. And real quick,
- 7:59verification isn't Claude reading its
- 8:00own work and saying, "Looks good to me."
- 8:02It needs some kind of evidence outside
- 8:04of that first draft, whether whether a
- 8:05screenshot, a test result, [music] a
- 8:06source, a reference example, or like I
- 8:08said, feedback from a few different
- 8:09agent perspectives. And you can add
- 8:11something like this into almost any
- 8:12skill. Say something like, "Before
- 8:14returning the final output, define the
- 8:15acceptance criteria, create the first
- 8:16version, inspect it using the relevant
- 8:18verification method, fix every issue you
- 8:20find, and then run another pass. Return
- 8:22the output only after it meets the
- 8:23criteria with a short summary of what
- 8:25you checked. And if something can't be
- 8:26verified, then tell me exactly what
- 8:28remains." It's even better if there's
- 8:29some sort of objective success metric
- 8:32that you can set and just have the
- 8:33agents keep working until it hits that
- 8:35objectively. But anyways, now the first
- 8:36output is basically an internal draft.
- 8:38Claude reviews it, catches the obvious
- 8:40problems, and then improves it before
- 8:41you see any of it, before you waste any
- 8:43of your time and attention on it.
- 8:45Because you're ultimately still going to
- 8:46be the final judge, especially when
- 8:47there's things like taste or strategy or
- 8:49business judgment involved in the
- 8:50process. The goal though is just to stop
- 8:52spending your time catching problems
- 8:54that the AI could have caught on its
- 8:55own. Your first look should not be the
- 8:57agent's first look. It should be the
- 8:58agent's, you know, fourth or fifth or
- 9:00maybe even sixth look. So now you know
- 9:01what Anthropic engineers are actually
- 9:03building. They save proven code instead
- 9:05of rewriting it. They give every skill a
- 9:06precise description so Claude loads only
- 9:08the correct one. They turn corrections
- 9:10into durable instructions that keep
- 9:11improving, and they verify the work
- 9:13before it ever reaches you. That's how
- 9:15you take a general purpose agent and you
- 9:16teach it how you work specifically. And
- 9:18if you want to see how I organize the
- 9:19agents, the context, connections,
- 9:21capabilities, and cadence around all of
- 9:22this, then I'll put my full AI operating
- 9:24system course on screen right up here. I
- 9:27also have free courses, templates, and
- 9:28resources inside my free community that
- 9:30will help you build agents and workflows
- 9:32from scratch. You can join with the link
- 9:33in the description, and if you want to
- 9:34go deeper [music] on turning these
- 9:35skills into a career, then you can check
- 9:37out my plus community. But anyways, that
- 9:38is going to do it for this one. So if
- 9:40you guys enjoyed it or you learned
- 9:40something new, please give it a like, it
- 9:42helps me out a ton. And as always, I
- 9:43appreciate you guys making it to the end
- 9:44of the video. I'll see you all in the
- 9:45next one.
- 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.