HTML is the new Markdown: How Anthropic engineers are building with Claude Code | Thariq Shihipar

Markdown became a really popular way of interacting with agents, but the plans are so long, I honestly have stopped reading them. And this is honestly a mistake. I think that you still need to be really in the loop. Plans matter. PRDs matter. Spec matters. When you say, okay, Claude can run for 8 hours, what you're really saying is Claude can spend 500 bucks. All of us are becoming these compute allocators now, right? And so you have to decide what is worthwhile spending the compute on. People ask me all the time, Claire, you said product management is dead. What's next? And I'm going to say, you're a compute allocator, babe. That's the job now. You're still doing the same thing, though. You're writing documents to decide whether or not something else should do work in the shape of that work. Okay, so you've convinced me HTML is the future. And I like how you said this. It's not that it is necessarily harder or easier for the agents to read. They're very smart. They can read all sorts of code. But in fact, what you're finding is that HTML makes it easier for you to engage with the content, which then uplevels the quality overall because you're not, your eyes aren't crossing, looking at a bunch of raw Markdown being like, whatever, it's probably good. Instead, you're actually getting pulled into the spec or the document or the plan and then interacting it with a way that upgrades the quality and then you can ultimately build something better. That's right. Okay, so you're building something with the agents so the agent can manage you. You know, I'm not sure if manage me is the right word exactly, but you know, I just, I care a lot about being in sync with the agent. This is sort of like the features that I built in Cloud Code have been like that, you know, like how can I get to know you better? So, yeah. Okay, great. Well, we have Cloud Code up. So let's walk through how that works. Yeah, so I did like a little bit before we started. And so I like to talk with Claude just as a human, you know, and like I always start with brainstorming. It's so much easier to brainstorm once you, like, you know, once you have a partner. So I was literally like, look, I'm on a Clearvo podcast. I want to do a demo, you know, and can you brainstorm me some ideas in an HTML file? Right. And this is literally the prompt I gave it. Right. Like there's not, it's not complicated. And so here you can see the eight visual demos that it made for me. And it has these little mock-ups as well. Right. So like purity to working prototype. Right. Like it searched you up. Right. You saw that with a web search. Right. Whiteboard sketch to working UI, which I thought was really cool. This is such a cute, like little thing. It is extremely cute. And I, it's what's really funny is just this morning a ChatPRD user messaged me and they're like, I love the mockups in ChatPRD. And I'm like, what in the world, what are you, what are you talking about? Because I have something very similar to this in code review right now and haven't shipped it. And I'm like, did I, I was like, did I act to Claude accidentally do this? And it was that like cute little ASCII, you know, wireframe. So this is definitely the dream, but not even, but now you're telling me I'm going to build it. So, so it's giving you basically, instead of saying, here's a Markdown document of kind of like what you should talk to Claire about, some descriptions of things you could do. Instead, it was like, what's the best way to convey this information so you can actually engage with it and pick something. And it used HTML to make this visual guide of a potential agenda or a set of demos. And you just get like a much richer expression. Yeah, exactly. Like, I think like another, like for brainstorming, one of my like sort of rules of thumb is that I'm not going to read a longer output than the screen on Cloud Code, you know. So like if I, if you gave me eight ideas, I'm just not going to see all of them. And, but with HTML, I'm definitely, I scrolled through all of these, you know, and yeah, the, the, the diagrams just make it so much more evocative for me to like, sort of understand what's happening, right. The slash command starter pack, five code of feature flag dashboard. Yeah. PRD diet. And the one I ended up liking the most was a CSV to interactive dashboard. We love a dashboard. Yeah. I used to say when I was in enterprise, I guess I still am in enterprise software, dashboards equals dollars. So I like this one. Yes. Okay. So you use Cloud Code. You said brainstorm, but brainstorm in HTML. Give me a couple of things that I can talk about. It gave you eight ideas, including visuals and this lovely, like why her, what the visual is. And then the, I liked the risk. It's like it could go sideways as all good demos. Yeah. Can. And so you're going to pick one and then you're going to show us how you pull this through to a full plan on this idea. That's right. Yeah. So I think the, what I like about HTML is like really Cloud really understands this. And so my next prompt here was really like, okay I asked it to make some mock-ups in the follow-up prompt. And then I was like, I asked it to interview me about number eight. Right. And so this is something that like, you know, similar to specs and PRDs, right. Like finding out my unknown unknowns, what do I want it to do? I answered a bunch of questions and now I'm like, okay, create an HTML file as a plan that helps me visualize what the implementation plan is. Include excerpts, mock-ups, code, whatever is needed to give me like maximum context. Right. And so then it made me this HTML file here. Yeah. You can see now this is, this is the plan, but it's, it's purely in HTML. It's got like, it started scripting out the podcast itself, which, you know, maybe I didn't need all of that, but like, you know, we're making a skill. And so, you know, fleshed out the, the file system. It gave me like an excerpt of the skill.md. Um, it put together like a, a mood board as well. Some example components, um, some of the logic here. Um, yeah. Insights and templates, helper scripts, right. And like helping me get a sense of like what's the important things for me to know here. Um, yeah. And this is, this is something that like, I will actually read, you know. Yeah. And I want to go back to Claude Code really quickly if you don't mind, which is, you know, people are gonna ask, how did it know exactly what to put in the spec? And I just, I want to go back to, your prompt was very simple and it's so funny. I've done. I don't know, 75 of these How I AI episodes and they get incredible outputs. And everybody's prompts like make the thing. Hopefully it's good. Yeah. Kind of nice. And so I love that this prompt is literally just create an HTML file with a plan. Help me visualize. You misspelled excerpts. I did, yeah. And you're like excerpts, mockups, code, et cetera, whatever is, is needed. And so I do want to encourage people, you know, don't stress so much about what should go in the thing. And in fact, it might change initiative to initiative. It might be slightly different to engage you with the work, but like identifying what you want to get and then letting Claude, letting the model do what it needs to do. We'll do a very high quality job. Yeah. I think with prompting, it's like this fine balance of like, I think you want to give enough information that you get what you want, but you don't want to over constrain Claude, you know. And so sometimes when I see people with a lot of like overbuilt skills, kind of like, you know, you're an expert planner or something, right? Like that is usually like outsourcing too much and constraining it. But in this case, for example, like I really did want it to make sure it gave me code excerpts and I wasn't sure if I did, uh, whether it would do that, you know. But this was really important. But then I always need to give Claude an out, you know. I always needed to be like, okay, like you asked me for this, but you know, like there's something else I want to give you. And so whatever is needed to give me maximum context is like my way of saying like, Hey, Claude, like, I trust you here. I want to just like be in the loop with you. Yeah. I love what you say, which is I trust you because my new ending prompt is not make no mistakes. Love make no mistakes, but that's not it. I, I literally like, I believe in you and trust you. Exclamation, exclamation, exclamation. I'm like, truly. I know you are capable and I believe in you to make these decisions. And so I think leaving that open-ended sort of like whatever you got to do, I trust your judgment. Um, at least it makes me feel like I get better outcomes. Yeah. Yeah. I mean, I've loved the like recent twist on this where it's like make mistakes, Claude, you know, like fall in love The types, like what's going in, and then this is how you would test that you did it correctly. And with those two bookends, you can like everything in the middle is kind of gravy. So that's what I think about. I think that's right. We could have a whole podcast on testing, I think. Let's do it. Okay, round two. We're gonna get it scheduled. Yes, yes, yes. Yeah, my tagline there is like, verification is not testing. And so I think there's a lot of like, there used to be like unit tests and things like that. But now, yeah, verification can be like a rubric, like you're with managed agents and outcomes. It can be like, have Claude record a video of what it did for you, you know? So there's a lot of depth there. Yeah, I keep a set of like synthetic data and I like run a CLI through this synthetic data. Because I'm like, these are all the things that have broken in the past. And if you get better at like resolving these broken things, then we have moved forward. So I think there's just a lot of interesting verification and testing mechanisms you can do now. We are gonna... Part two. Pressure him to do it in the comments, please, please tell us. This episode is brought to you by Persona. You're learning to build with AI, but there's an important question you need to ask. Who is actually using your product? Is it a legitimate user, a bot, or a fraudster? Brex, Figma, Etsy, and Twilio trust Persona to answer that question. With Persona's identity verification platform, you can create branded experiences, automate fraud prevention, and know who is human online. That makes it easy to give good users an experience that makes them feel welcome and to stop bad actors from causing damage. And for those of you building in the AI agent space, Persona helps you verify the identities of people, businesses, and developers behind agents. It's how companies like Lithic and Skyfire are pushing the frontier of agentic commerce. Learn more at withpersona.com. Okay, so you have built this, but here's the objection I'm gonna get, which is Markdown is accessible, right? I can like go into Markdown, type in it, and make edits. And so I think that is one reason is it has been so popular to bridge this gap between machine writable and readable, human writable and readable at a very low level of sophistication, right? As soon as you understand, okay, these like hash signs mean headers, you're good to go on Markdown. How, like, I want to fix this. How do I touch this? How do I edit this? Yeah, yeah. So I think that this is, like, a great point, right? And I think one thing I felt with Markdown was that, one, because I've stopped reading them, I had stopped editing them as well, you know? And so I would end up asking Claude to edit them. And so, like, that is, like, the most basic form is just to be like, hey, Claude, I didn't like this part of the plan. Can you edit it? But let's say you want to get really in the loop, right, and like really get in depth with it. Claude can also do that for you. So the next prompt I had, and I forgot if it was here, it was here. Okay, I want to create an editable HTML artifact to help me define the decision rules. So these are the rules that it's defined here on, like, okay, how do you take data and turn it into, you know, a visualization? And I think some of these are kind of arbitrary. And so I asked it, like, you know, creating an HTML artifact, I don't like the ones we have right now. Make this a custom UI that helps me with structure, but gives me flexibility. Design the ideal interface for this problem. Yeah. I really wasn't quite sure what it would give me. And this is one of the fun parts of HTML too. It's just like, I just want to see what Claude, like, cooks up here. And yeah, this was what it gave me, right? So it's like a my own beautiful custom interface. I can sort of like, you know, edit any of these fields. I can, you know, like hide them. I can copy. I can, you know, add new fields here. And it gives me a Markdown to copy back. And so once I'm like, okay, I have this, I can copy it back into the output. Okay, I want to pause because people are going to totally miss what you just did. So I'm going to repeat it, which is you have this HTML plan and there's a section in the HTML plan that is a pretty like specific table of rendering and visualization rules per data type that you could predict would be in a CSV. And you're like, I don't like it. And instead of going back into Claude code and being like, I don't like it. Let's go back and forth and edit it in like the terminal. You said there's probably a way for me to interact with this particular problem that's ideal from a user perspective. So basically build a throwaway UI for this very, it's like, this is not even personal software. This is like sub, it's like micro software on top of micro software, which is like, I've made this very personalized plan and then I'm taking a module in the personalized plan and zooming into it using a very custom UI that's going to engage me with the content to get to a higher quality. I also like that. It's like kind of gamified. It's like very consumer. Yeah, yeah. Very consumer. And you said in the prompt, give me the ideal UI for this to like help me engage with this. You built this, then you get the data, right. And then you're just going to bop it back into the file. Yeah, exactly. So fun. Is this how you're building now? Actually it is. Yeah. And do you have any challenges with, like, how are you passing this around from a collaboration perspective? Or is it just like, this is the way a single threaded product or engineering leader can get something done. It's you're engaging with yourself and with the model. And so you feel like you can own things full stack or do you hit friction points with collaboration? Like when somebody needs to give input on this. Or you want input. Sure. Yeah, yeah. I mean, I think on the scale of an implementation plan, it's way better. Right. And I think that this is because you just like can upload it to like, you know, whatever, AWS or something. And then you just share the link around. And so definitely the like likelihood of like, I don't know, Kat or Boris like reading this is like a hundred times better. Right. And so I think that really helps me, like, present this. I also just, you know, somewhat related, I use it a lot in like collaborating overall. So for example, you know, I report to Kat and so like every week I send her a weekly status update in HTML of everything I've done. I get Claude to read my Slack and just like create this message. And, and like she actually gets to read it and I don't have to spend that much time on it, you know. Oh my gosh. New comp, new internal competition is showing up. It's not just who is building the best product. It's who's building the best product that goes into building the best product. And like who is building the best product to represent themselves to the manager. But I mean, I think why you do that is not artificially for fun. It is that it's just a much more effective way to communicate across a company is with content that is engaging and at the right level of detail and consumable. And we're all pretty good at reading websites. Yeah, exactly. I think this is like, when I think of like abundance, you know, and like, you know, we talk about like Javon's law for software, like, oh, like you software gets cheaper. What do you do? I'd say like the amount of tokens I produce that go into production code, like extremely small. It's like 1% or something, you know, but like I'm generating so many more tokens like this, like my dashboards, my customer interfaces, like really trying to get a sense of like, what do I want to do? And yeah, it's like I have everything I'm interacting with is so beautiful. And I think my hope is that it also like translates into what I produce in the end. Right. That it's like more in the loop. It's more beautiful. It's more like, you know, like what me and Claude working together. Yeah. And I like this because I've been in the, in the product game for quite some, some time, many decades. And people used to get so wrapped around the axle on like, what's our source of truth for specs and what's our source of truth for PRDs. And you know, is all this information in some centralized place that we can all access it? And is it all in the same format? Is it all on the same template? And there were these arbitrary rules because creating these content was relatively expensive. Consuming it was certainly expensive. Finding it was really hard. And I think when all of that cost goes to like functionally zero, although we're all paying her. I call them Claude chips. We're paying our Claude chips. But you can kind of put stuff wherever in whatever format, because we know these models are very good at using tools to discover the context that they need. And so I do think there's this fun moment where you're really up, you like up-level the things that you should care about, which is like, what is the content of the plan? Is it a good idea? Do we feel like it's going to be So, so co. I thought it would be on brand. Okay, second thing, we were at this amazing event. What is the thing that you're most excited about or that you saw or heard today? I think obviously we had a big announcement at the start of the day, our partnership with SpaceX that bringing the more compute online. I think, yeah, I'm excited for, you know, we also said we are thinking about orbital data centers, and that's just- I love it. Yeah, you know, incredibly sci-fi, but could, you know, it could actually happen. So, yeah. I know, we were watching this moon mission with my kids who are like kind of elementary school, and I'm like, would you want to work in the moon mine? Would you want to work in the moon mines? Because I think it's coming. Yeah, orbital, orbital. That's, that'll be next year's demo, right? You and I will come back and we'll do this, we'll do HTML, we'll do testing, and then Lunar Data Modules. That's right. Perfect. Okay, and then my last question, very important. I love that you just talk to Claude like a person when Claude is not listening, not giving you what you want. Yeah. What's your prompting technique? Do you yell? No one in Anthropic yells. That has been my experience so far. So you'd be the first to admit it. Yeah, yeah. No, I definitely don't. I think that like, there are a couple things here. I, I do sometimes message people and I'm like, hey, like, seems like you have a bug. Can you send me your transcript? And they're like, It's not the best side of me, yeah, you know? I think that like, yeah, I don't, I don't yell like Claude. I think that like, we've also done some interesting research recently about like emotions and in Claude and just sort of like, this idea that like, once you, when you say things with a certain emotional charge, it also like activates different features inside of Claude code. I don't think anyone's done this like A-B test of like which, like, you know, if you're mean to Claude, does it better than without it or not? But I'm just like, let's err on the side of like, you know, just, not, or like, what's the thing I'd prefer to exist, you know? And I'd prefer if you're like nice and, you know, friendly to Claude that you get better output. So yeah. All I've seen is if you sort of border on stern to any of these models, their reasoning gets really sad. It's like, oh, the user is right to be so disappointed in me. I'm like, oh, I don't want, I don't wanna read that. I don't wanna see that. Yeah. Thinking traces are tough. I usually give the model some privacy. I'm like, I'm not gonna read the thinking traces. Like none of your business. Yeah. I had somebody else on, I feel like it was Hillary, who was like, just like an employee. How you get your work done is none of my business. I don't even wanna know. We just collapse those things. Well, this has been so fun. Thank you for showing us the way. Where can we find you and how can we be helpful? Extremely online at X. Yeah, I'm at TRQ212. And yeah, just tag me if you have, you know, anything with Claude Code. I'm happy to help. Perfect. I am living proof. This man is happy to help. Well, thank you for joining How I AI. Thank you. Thanks for having me. Thanks so much for watching. If you enjoyed the show, please like and subscribe here on YouTube or even better, leave us a comment with your thoughts. You can also find this podcast on Apple Podcasts, Spotify, or your favorite podcast app. Please consider leaving us a rating and review, which will help others find the show. You can see all our episodes and learn more about the show at howiaiPod.com. See you next time.