[한글자막] Anthropic 엔지니어 | 저희는 Claude Code를 이렇게 사용합니다

追加: 2026-06-01

[00:00:02][music] [music] [music] >> Hello, hello, hello. Welcome all. Thank you for joining the workshop.

[00:00:26]We have a pretty, pretty full house here today and really pleased to see that. Um quick show of hands who here loves Claude code?

[00:00:38]Okay.

[00:00:39]>> [laughter] >> You're in the right place. Yeah. Yeah.

[00:00:42]Lovely to meet you all. My name is Arno.

[00:00:44]I'm a member of the Applied AI team. I'm an architect there. I'm here today to tell you how we Claude code at Anthropic do a workshop and uh you can code along. You'll get some credits. I think there'll be a QR code in just a moment, so make sure to grab that. Uh there'll be people floating around that might support you. So if you have technical difficulties, we can we can help you out. Um Anybody here in this room who hasn't used Claude code before?

[00:01:13]Okay, great. Good. Good. Cuz this is uh there's more to do. There's more to more more to find out. I'm happy you're here.

[00:01:20]Let's get started. So let me get my clicker.

[00:01:24]Ah, yes. Excellent. Yes, please, please, please grab the QR code set yourselves up.

[00:01:31]There'll be a repo here as well that you can clone.

[00:01:34]Uh there are three phases in it we're going to work through.

[00:01:37]There's an interesting verification setup that you can work through alongside me uh and it's quite detailed, so you probably will want to investigate it a bit later yourselves.

[00:01:47]What we're covering today is based on uh this version of the talk that Tarik gave uh in San Francisco just about a week and a half ago. Who here follows Tarik on on Twitter?

[00:02:00]Fantastic. Great. Yeah. Yeah. So, he published that as a blog post called The Unreasonable Effectiveness of HTML files. And um he's basically pitching that moving on from markdown files, we're going to be using HTML instead. We're showing some of that today. Like I said, there's a repo to go along.

[00:02:17]Uh enjoy that along the way and later on in the in the week as well.

[00:02:23]Where are we now? Where's this all going?

[00:02:26]I think everybody probably notices that agents are becoming more and more capable. Why is that happening? That's happening because the models are becoming more and more capable. And so if the models become more capable, it means agents can run for a longer and you can give them more and more complex tasks. But that also means that we have to change our way of working. You have to change our habits.

[00:02:46]That's part of what today is about. How can you change the way that you're working with Claude code to get more out of it?

[00:02:55]If you're going to let your agent run for a longer period of time, then you can burn through a lot of tokens if it does the wrong thing.

[00:03:04]And you want to avoid that ideally at the beginning.

[00:03:07]And so that's where the idea came from to to to push more and front-load more of the verification that the human would do in the spec into an HTML file because it's a more rich and more um more human ergonomic way of engaging with the content that your agent is going to be building for you.

[00:03:28]And we'll see how that works.

[00:03:32]There are a bunch of things around agents that are worth reading. We have a lot of engineering blogs on our web page.

[00:03:39]Um we also have the the standard blogs.

[00:03:41]You should check all of those out.

[00:03:42]There'll be a lot of information there.

[00:03:44]There's a lot of interesting information about harnesses, about long-running agents. All of these things are important. Today we'll focus on three levels. Um and there'll be some something a bit more basic, something a bit more next level, and then something that's quite interesting and that will be a bit more in-depth.

[00:03:59]The first thing is the more capable the models get, the more you should try to resist constraining them. Who here, by the way, has heard of the bitter lesson by Richard Sutton? Great. Fantastic. This is This is great. So, I'll I'll I'll For those who haven't raised their hands, uh, Richard Sutton is basically the the father of reinforcement If you were reading a book about reinforcement learning, it's probably authored by him.

[00:04:24]And, um, his idea was basically that, you know, you could spend all your time trying to with your human capabilities hard code up front and constrain the system, but in in the end, pouring more data and more compute at it ends up getting more capability than anything that you could have come up with. And there's a similar analogy here, right? The The models are becoming more and more capable.

[00:04:46]And so, you should accept that the model is probably better at extracting requirements from you than you are at defining your requirements. The The requirements are latent within you, just like when you talk to your users. Your users, they have an idea of They know it when they see it, but they're often not very good at articulating what they need. And likewise, you probably know what you want when you see it, but Claude is likely better at extracting what you want and what you need from you than you are in specifying it to Claude. That's another direction to take this in.

[00:05:19]So, we'll talk about that, removing ambiguity, letting Claude prompt you and interview you in prompting.

[00:05:26]Then, how to then understand and plan.

[00:05:29]We used to do this all with with Markdown files. We still do it with Markdown files. Colleague of mine once said, "The Markdown file is the lingua franca of the AI native software development life cycle." That was pretty poetic.

[00:05:40]Um, but it seems like that format is a bit constrained. It's getting too long with a lot of lines of markdown files to read.

[00:05:49]We can condense more information into HTML files. And that's what we're going to do today.

[00:05:54]And then how to verify.

[00:05:56]Not test actually, but just to verify in this context. Um how to make verification native to the thing itself, so that the agent can drive it alongside a human. Or eventually not headlessly as well.

[00:06:08]That's where this is all going as well, right? The agents are going to be doing more and more of this natively. And how can you set the artifacts that you produce up to natively be testable and verifiable in the way that you need.

[00:06:22]So, we'll do an example.

[00:06:24]By the way, has everybody had a chance to get grab that QR code and set themselves up? Very good. There'll be a link to the repo eventually as well. I think that comes in a bit.

[00:06:34]Uh let's say we want to do a bill splitting app.

[00:06:36]Uh very simple, um you know, you want to you go out with friends, you want to find out who owes what. Yeah.

[00:06:46]Um let's let's Claude interview us around doing this. So, in this case, I'm actually going to Yeah, before I do that, give you an example of how you would do this and how you could do this better.

[00:07:01]Yeah?

[00:07:02]What's good prompting? What's bad prompting?

[00:07:07]Bad prompting is when you say just make it better. And a lot of people that I watch using Claude code just type make it better. Um make no mistakes. Yeah.

[00:07:17]Uh that's not good prompting. Um you want to you want to you want to encourage Claude to extract from you specific details. So, you give give the domains. Don't over specify the outcome, but specify the areas that you are interested in. Right? That's what makes this good prompt on the side different and better. Um We're making you know, focusing on the audience, for example, or uh suggesting an open-ended way to answer the question as opposed to predefining it up front. And then that will prompt Claude to iteratively interview.

[00:07:51]All right. So, I've got a Claude open here, two different Claude's.

[00:07:56]Who here has used fast mode?

[00:07:59]Okay, not that many people use fast mode. That's why set it up here.

[00:08:02]Um and who here uses auto mode?

[00:08:08]Oh, I'm very happy that you all use auto mode. You need to be using auto mode. If you're not using auto mode, you need to be using auto mode. It makes it so much easier.

[00:08:15]Um yeah, use auto mode. Good. Who here is setting their effort parameter?

[00:08:21]Good. Good. Our recommendation is X high, but you can also set max effort. I mean, in my case I think I kept it at X high for this.

[00:08:28]Um yeah, yeah, yeah.

[00:08:32]So, just to touch on how those look, right? Like we've got forward slash effort, which is the effort parameter, forward slash fast, which is the fast mode, which is I've I've turned it on. And then we have the auto mode, which we cycle into like this shift tab, yeah. Auto mode is the best. So, I'll copy paste my my You all, when you have the repo, will see um one of these prompts is in there. So, I'll just copy paste the prompt in here on the the bill splitting app.

[00:09:01]So, okay. Claude's going to ask me uh you tap through these, right? Like you'll tap through these like this.

[00:09:07]Uh and in my case, I want to uh I want it just for friends.

[00:09:14]And I wanted uh Is there a secondary audience? No, no secondary audience. And we'll leave that as that.

[00:09:26]The key thing here in the prompt is that you want to use You want to prompt Claude to use the ask user question tool that you saw me use in the prompt earlier before, let me submit those answers.

[00:09:36]And then it'll write that spec, right?

[00:09:38]So, you you saw me in my prompt explicitly referring to the ask you the question tool. That's what triggers this workflow. And, you know, depending on how well you specify that prompt, you'll get better outcomes. So, it'll generate a spec. We could then turn that into a bunch of different ways of of of building the app.

[00:09:59]It's going to take a while, so we'll go back to my slides. Could I go back to the slides, please?

[00:10:04]Great. So, let's say we have a plan.

[00:10:06]Generates a plan. We've answered some questions. It it it's getting better at extracting turn by turn from me what I actually want without me having to upfront verbalize everything myself.

[00:10:21]Then I want to be able to check, is it what I want?

[00:10:25]Well, an HTML file is more dense.

[00:10:28]Much more information dense. Much more ergonomic for you to understand what the thing is going to look like. You can even use a screenshot with it. You can use a Playwright MCP later on for that as well, right? You can interact with it much more richly than you could with a very long markdown file. And especially if the markdown files get, you know, they get more than about 200 lines long, it's unlikely you're going to read it.

[00:10:50]And certainly unlikely that your colleagues are going to read them.

[00:10:53]Before we started here, I had Claude with Opus 4.7 generate a few examples for me of what this could look like with my bill splitting app. And we'll look at those now.

[00:11:04]We have them here.

[00:11:07]The prompt I used is in the repo that you'll have access to. I asked Claude, give me a few different directions, four different design directions, explore them, generate them as HTML, and let me explain it explore them across each other. And they came up with these different ones. So, one sort of brutalist, one Tokyo fintech.

[00:11:24]We'll we'll click them all one by one.

[00:11:26]So, this is what this one would look like.

[00:11:29]Uh and then maybe this one this one might look like this.

[00:11:32]Completely different aesthetic, right?

[00:11:34]And I click around and And this is much better for me to give feedback to Claude on than it would be to just try to infer from a markdown file what the thing is going to look like. I could go in and, like I said, take screenshots, feed them back into Claude. Who here regularly takes screenshots to give information to Claude? Very good. You should be doing that. That's very good. Um especially when you're when you're doing front end, it's really hard to articulate like the thing is slightly off or there's a misalignment here and it's like you'll find that you run into the limits of what you can express and it's actually easier for especially Opus 4.7, which has a much better vision model than before, um to extract from you what the problem is and proactively.

[00:12:16]Great.

[00:12:18]So, a few examples. Tariq has a lot more in his repo, which you'll find in the repo as well online. So, so far, we've covered letting Claude extract information from you interactively as an interviewer because the longer you let a an agent run, the more important it is that the spec is comprehensive, the less likely it is that you will be able to upfront define everything and it's better for you to iterate with Claude in that manner.

[00:12:52]Then, what's a better and more efficient and ergonomic form factor?

[00:12:56]Um it would be the HTML file.

[00:13:01]And now, the more important part from all of this is how to verify what Claude has done and how to make it agent native. That's what we're going to cover with the repo.

[00:13:14]I think we have a I think there'll be one slide where you'll get to see the actual slides to to engage with uh the actual URLs to engage with. So, what you want to do is make it part of the artifact. And that's what we're going to cover today. In fact, there's part of what we're doing today is we're going to we're going to use Storybook Fixtures, a testing library, sort of data emissions, and attributes, and Playwright MSTP for a React uh for a React app. Um with with components that you will have seen before, but remixed in a way to make it easy for Claude as an agent to extract the data contracts in the DOM, and then run verification, and then even record the verification, which is how the Claude code team that Tariq is on runs this, and that's what the demo that he built refers to, right? Like it'll be a generation of verification steps that get run through, they get a recording, and that's then put on S3 or I share it somewhere else with colleagues, and that's how you turn these verification steps into something that is generated by the agents, and then that you have fewer and fewer touchpoints with.

[00:14:31]That's something that you can replicate especially from that repo.

[00:14:35]The way to do that is to sort of sensibly modularize, uh but that becomes really clear when we look at the actual the actual thing.

[00:14:43]So, I encourage you to go to that link.

[00:14:50]There'll be a repo.

[00:14:51]This is the repo under csw you see workshops, Claude with Code Workshops, and there's the one for today's session, which is how we Claude code.

[00:15:02]What's covered in there is phase one, which we just covered, right? The simple prompt to say, "Hey, I'm writing a bill splitting app. Interrogate me on the requirements so that I don't have to miss any in my upfront description to Claude." The second one is generating four different HTML design directions that we could explore uh for me to then feedback on or for for me to decide. And then thirdly, we have a verification framework here.

[00:15:28]Uh And it's in a different context. It's not to do with the bill splitting app. It's regarding a to-do app, a small to-do app written in React.

[00:15:40]And we're going to show how we build verification entirely into that along the way. That's described in detail in the readme for phase three and then also in the verification detail as well, which is here. So, if you want a deep dive on how this is set up, you can read this.

[00:16:03]And it's all provided in the repo. It's pretty cool. What is it focused on?

[00:16:08]We've written a a small to-do list app here.

[00:16:16]I can add an item, test.

[00:16:21]It's hard to see. There we go. Very good. And then I could tick it off and drop it as well.

[00:16:27]And I could clear the finished items.

[00:16:29]And so, many things are happening here in terms of the state, and we'd like to verify that it's all working alongside the tests that are already defined.

[00:16:37]Uh but these verification steps, we'd like them to be driven by an agent.

[00:16:41]There are three ways that we're going to do this. So, we're going to do this in a human-readable way, um but then that same way of doing it in a human-readable dashboard, we're going to verify uh in an agent-first way, uh and let Cloud do that separately. And then there's a separate way in the in the repo as well.

[00:16:56]You can just run um run run verify, and then it'll just run the test matrix there. You'll get a slightly different result if you do it from the repo versus here, but the the principle is the same. So, human-readable, agent-driven if you're doing it from from Cloud Code or somewhere else, uh and then also just generally headless like you might do it in CI.

[00:17:17]We'll have a look at it in a bit more depth. We'll go here into the app and we'll actually look at There we go.

[00:17:28]We'll look at elements.

[00:17:31]We have here the actual emitting so the the emitting different data data aspects, the data verify unit, the total done and active.

[00:17:40]Um the component itself here is publishing its state to the DOM. Uh and so if I change my if I change my state here, if I make I might add test again.

[00:17:51]See that updates. And then I drop it. It updates again.

[00:17:56]So, this is what the agent can read later as opposed to having to scrape the DOM. You can just if you publish the state here separately from the React internals, you'd be able to run the verification independently of what whatever the state of the app is.

[00:18:11]Um and that's how this is set up to work. Uh and then you can run this further down the line as well.

[00:18:20]Each one of them gets this, right? If you look at um if I if I pull up the That's a bit hard to read. Let's do that here. Um Ev- Every Every component gets one of these, right? There'll be schemas, there'll be fixtures, um the known states, and then the invariants, uh There are particular ones here that that always have to hold I always have to hold. You'll test them with probes.

[00:18:44]And we'll do one example which we've hardcoded which we've hardcoded an example that will fail the verification and we'll let the human verified dashboard catch up, but then we'll also let the agentic way of doing it catch up as well. And then we can let Cloud Code find it and diagnose it for us.

[00:19:03]So, that same approach also works here.

[00:19:07]So, I've got the dashboard defined here.

[00:19:11]And what we see here is, there you go.

[00:19:13]Um Like I said, we've got the different schemas, the the invariants that we've defined.

[00:19:18]And we can run these, right?

[00:19:19]Individually, we can see how they would execute here.

[00:19:23]Got an example here.

[00:19:26]Or also I could run all of them. In fact, I'm going to do that now.

[00:19:31]I run them all. One of them triggers artificially because we planted it before.

[00:19:37]Going to scroll down and find out what that is.

[00:19:40]And it's here.

[00:19:41]Uh we will actually look at how we can replicate this for an agent as well.

[00:19:50]And we would do that like this.

[00:19:51]Go here.

[00:19:56]Inspect.

[00:20:02]And we could run basically the manifest of all the different verification steps that are defined here as state in the DOM.

[00:20:11]We can get them in the same way that we have them here.

[00:20:14]We expand that. It's getting a little bit small.

[00:20:22]As an example like that.

[00:20:26]And then like that as well.

[00:20:31]Great.

[00:20:37]If we do that, then we can also actually run them uh I just specifically here.

[00:20:43]Uh I'll do it manually first.

[00:20:47]Let me replay uh replay them all.

[00:20:51]Close that.

[00:21:03]In this case, we're running this not to perform the verification, but but to provide the evidence of the verification. We can later on record this. We can record these as clips which would just be videos we capture. And then we would store them, share them with colleagues, put them on S3 or whatever it might be.

[00:21:26]Let me pause that here.

[00:21:30]There we go. So, here's our summary.

[00:21:32]We have one that's deliberately failed.

[00:21:34]I'll explain that in a second.

[00:21:36]But in each case you can also see the details of how this was done.

[00:21:45]The key thing here is that we want to um it's important to include the probes to to push off the happy path. And then a lot of this will be generated by Claude for Claude, right? So, there's there's a way of scaling this further.

[00:22:00]Um in the end, we'll go to the one that didn't work.

[00:22:03]So, the one where we hard-coded that the sums don't match. In this case, the the state doesn't match. Uh 3 + uh uh 3 + 4 does not equal 10 in this case.

[00:22:19]And we'll get the same result if we run verify all from from the Claude here as well.

[00:22:31]And we'll let Claude do that in a moment as well.

[00:22:35]There we go.

[00:22:49]Then it's a little bit hard to see, but there's pass pass pass pass pass and fail.

[00:22:54]So, there we go. Okay, good.

[00:23:03]We can change this, right? I can The The point here is that I'm I'm trying to show that the The idea is to get something agent native be read as the DOM contract here. Um that's what we established earlier that the state is managed here or it will not manage but at least um viewable to an agent that it can then run the verification end-to-end itself.

[00:23:25]And we can let Claude do that afterwards as well.

[00:23:29]In this case, we can also make a change that will break more. Let me see if I can break the chain. I'll break the contract but not the app.

[00:23:38]And I'll do that here.

[00:23:43]So, I could change for example here.

[00:23:50]Under Exactly. Under to-do app total stats, I'll delete that.

[00:23:54]I'll do undo that afterwards as well.

[00:23:59]And if we now rerun then all of these ones at the bottom here will fail.

[00:24:23]And we'll get the same when we run this again from here as well.

[00:24:28]Oops.

[00:24:42]There we go. All of these are failing.

[00:24:44]Not because we broke the app but because we broke the contract.

[00:24:48]That Claude can then natively verify.

[00:24:54]Excellent. Great.

[00:25:08]I want Claude to tell me what's going on with these ones. This Well, not the ones that I broke, but I'll undo the ones that I broke. But, I'd like it to tell me what's going on with this one. And so, let me correct the change that I made before.

[00:25:23]Control Z.

[00:25:26]Auto save.

[00:25:27]Rerun.

[00:25:30]And there we go.

[00:25:42]So, we've done it manually, and we've shown how we can match what the agent would see versus what we would do. But, we can let Claude run this headlessly itself as well. So, I'll do that, and I'll pull that open.

[00:25:52]So, let it run.

[00:25:54]This one's broken deliberately, and the rest works.

[00:26:09]Great.

[00:26:10]So, we have Claude running here.

[00:26:13]Uh I mentioned fast mode before, I mentioned auto mode, which is great. I mentioned forward slash goal.

[00:26:19]In this case, what we're going to do is let Opus 4.7 help us find out why that particular verification failed.

[00:26:32]I've already connected the Playwright MCP for this.

[00:26:35]Uh oh, and it's going to use that.

[00:26:38]And it's going to run. Ski bar got rejected. 4 + 3 does not equal 10. Uh yes. If you were to run bun verify, uh run bun verify, then you would actually pass the tests in this case, uh because we deliberately put the verification in wrong just to demonstrate within the the test matrix itself will actually pass.

[00:27:01]Um But the idea here is to separate out what you could do as a human versus then what you could do the end agent directly from the browser using basically these the commands that you saw. And then also what you could run headlessly directly from the CLI.

[00:27:19]If you wanted to do it like this as well, then you could record the outcome that we talked about.

[00:27:29]Yes. So here, I've in this particular setup, you've actually got the ability to to show how you would record these. Um the same delayed running that we've shown, you could run and record as evidence. Basically to show that it would work, um and you could store that.

[00:27:48]Uh and they could run like that.

[00:28:00]This is very common right now.

[00:28:02]Um the the cloud code team uh records basically all the code changes that they do like this. Um all the front end changes at least.

[00:28:10]Uh especially on the round there at the pace of the shipping that we have at the moment.

[00:28:16]Are people able to pull up the the repo and get the verification set up working?

[00:28:20]Yeah, very good. Very good.

[00:28:29]Hm?

[00:28:31]Yes, you can store them. I mean, like you can just put them in S3 or whatever or short share them with colleagues or um in our case we have a version we have a we have a internal This is more automated than what I'm showing here. Uh in our case we do record them. Um not certain for how long or in what context, but uh it's part of a regular cadence.

[00:28:50]And then you get the In fact, you get the I should have triggered that.

[00:28:56]Did I not see that?

[00:28:58]There we go. So, you can you have each clip.

[00:29:00]You could download all of them or download them individually, and then that's basically the bundle that that that proves the verification worked.

[00:29:09]Um So, I guess to bring it around to what we've covered so far and then where it's going.

[00:29:22]Three different surfaces, the the human surface, like the um and then the the agent first from the browser um mapping on the same way.

[00:29:32]And then you can also run it in CI with just run bun verify.

[00:29:37]The objective here at the end is to figure out how do you embed the verification into the artifact itself.

[00:29:48]There's more detail. I encourage you to check out the repo itself.

[00:29:51]Um and actually run a few examples yourself and run a few tests. Right, you can change the code, see what breaks, rerun it yourselves. Um it's quite detailed, and it's what Tarik and and team use in the cloud code team uh for their work already. This is a this is this came uh pretty quickly um just a week and a half ago into this demo. Uh um So, I encourage you to check this out.

[00:30:17]What's new really is the remixing and the the new arrangement of of of of of of primitives that you're already familiar with that you're already using just to make it available to the agent first.

[00:30:30]That concludes really what I wanted to cover.

[00:30:32]Um I encourage you to spend more time on the repo.

[00:30:36]There's great documentation there, and you can actually um get a lot out of it with Opus 4.7. Opus 4.7 works really well because it has the better vision model. That's where this really excels. Um if you use Sonnet, I I wouldn't recommend that. Let's uh try using Opus 4.7 for it.

[00:30:58]Um try using Fast mode for it. Fast mode is great, costs more, but it's great for iterating quickly on specs.

[00:31:05]People will sometimes ask, well, isn't a HTML spec um more token inefficient?

[00:31:13]And the answer tends to be no. Uh and the reason is that in the long term you iterate less if you have a good and rich HTML spec, even if on one-off instances you spend more tokens to generate it.

[00:31:26]So, um you can even try it with Fast mode. So, that would be it. Uh that's my recommendation to you.

[00:31:33]I enjoyed uh speaking to you and thank you for your attention.

[00:31:37]>> [applause and music] >> Woo!