Lenny's Podcast: Product | Career | Growth
01:02:52 · Claire Vo
[01:02:52] S1: Yeah. Again, going back to that gnarly feeling of product market fit is it has so many sharp edges. I won't sugarcoat it. It's a pain to set up. You got to feed and maintain your Claws. It is not hands-off, but the value is so high, I am willing to go through the pain. And I think any good product manager has felt this experience of launching a product and it's just hot garbage. It's buggy, it doesn't look great. And your biggest complaints from customers are not, "I don't think this is useful." It's that, "It's broken or doesn't work as good as I want." That's when you know you have product market fit. So I like the complaints of like, "It's buggy. It doesn't remember. I want it to do X, but it can't." That's not, not product market fit. That's a product that hasn't caught up to its product market fit. And so again, I think this is just a really interesting one for people who are our product thinkers. But going back to the actual problems I've experienced, I think you outlined them pretty well. One, I don't think anybody has really unlocked browser use. That is not just an OpenClaw thing. I think we look at ChatGPT Atlas, you look at Perplexity, Comet, you look at all of these kind of browser use. Claude has a browser use component. I don't think any of them are great. And the reason why they're not great is technically I think it's a complicated problem to solve, so I'm empathetic to that. And two, the open web has been so hardened against bots. And so the way websites are architected are actually anti-bot. And so there are all these hard walls where bots can't access sites. There are all these bot identifying mechanisms. For example, trying to browse X with a bot, there's like a lot of punishments being doled out when people are found out they're doing that. And so I just think the web is hostile to agents right now. And we're going to have to rethink what is the interface of the web to be more agent friendly because I think we skip ahead a couple years and the number one user of websites are going to be people's agents. And so we can either say not allowed, you're not allowed to use websites or think of a new way to open up the web. Practically what I have done is, you might laugh at this because we've been talking about it a lot. I read the docs.
01:05:19 · Claire Vo, Lenny Rachitsky
There is a browser use documents page on OpenClaw and one, it explains to you how the browser works. So it uses a browser profile dedicated only to your OpenClaw that it can open up on its own and run very similar to if you have multiple kind of Gmail accounts and you're switching between Chrome profiles, exact same concept there. You just get one for your OpenClaw. [01:05:43] S0: And this is in Chrome.
01:05:43 · Claire Vo
[01:05:43] S1: And this is in Chrome. And you can tell your OpenClaw to give itself a color. So when you see the browser open, like pink is for Poly and green is for Sage, and you can really identify, especially if you have multiple agents, which one's working on which window. And then I know where browser use works and where it doesn't. So first thing you should try to do if you're trying to do something is look for an API. So does this have an API key? Your life is a lot easier. If DoorDash had an API, maybe they will have one, your life is a lot easier because then you can bypass the web stuff entirely. Then the second thing is, okay, if it doesn't have an API key, can it browse the website? And some things, and you'd be surprised which ones work and which ones don't. And I'll give a very specific example. I had Howie, the How I AI podcast producer open up YouTube Studio and look for comments that I should personally reply to. And so it was literally going into the comments tab, scrolling through, finding ones. It liked ones for me that I said that I liked and helped me find comments that I should reply to. And I couldn't get it through API. I couldn't do any... So I just gave it the browser. It was very slow, but it worked. But then I tried to have it do the same thing where I wanted to queue up some shorts in Instagram through Buffer, which is a social media platform. And the Buffer page was much simpler and it just couldn't figure it out. And so a lot of this is just trial and error. Will it work? Will it not? And if it can't solve a problem, I think where I've come and I would advise people to do is just walk away from expecting it to solve that problem and give it another problem to solve. So if you're saying it can't order DoorDash for me, maybe that's just not a problem for you, but can it meal plan for you? Yeah. Or you could say every day at 11 o'clock, I'm tempted to order DoorDash because I don't want to make myself lunch. So at 10:30, remind me of a few lunches that I really like to eat at home so I don't order DoorDash. So again, find out what the problem behind the problem is and see if OpenClaw can help you solve that.
01:08:09 · Lenny Rachitsky, Claire Vo
[01:08:09] S0: I love that. That browser limitation might lead to me eating better and saving money. [01:08:15] S1: We can hope. [01:08:16] S0: You mentioned at one point Exa. Talk about just where that plugs in, which is basically like a headless browser thing. [01:08:22] S1: Yeah. So part of what your OpenClaw can do is access the web. It can do that through a browser, so it can open up Chrome and search and all that stuff. But as we said, the web is a little hostile. Or there are these web search APIs that are now available. Brave is, I think, the one that ships out of the box with OpenClaw. And think of them as programmatic Googles, which is instead of going to a website and searching in a search bar, it sends an API request and returns search results and websites for you. So Brave is the one that ships with OpenClaw. I use Exa just because I already have an Exa account. You use Perplexity. Again, it's just a way to give your OpenClaw access to the web when it doesn't have fingers and can't use a browser very well. [01:09:10] S0: One of the other limitations I've run into is it's not just that I couldn't use the browser, it's that it disconnects from Chrome a lot and it's just like, okay, I can't do anything. I know they've been shipping a lot of updates there, so I think that's actually getting better. And I think that's just like, if things aren't working now, they're just only going to get better because now- [01:09:10] S1: Oh, 100%.
01:09:10 · Claire Vo, Lenny Rachitsky
[01:09:10] S1: Oh, 100%. [01:09:25] S0: This project is resourced and yeah, people are working on it. Okay. What about the memory piece? A lot of people complain it's forgetting stuff. You mentioned that. Do you have any advice for hardening the memory or just treating it a different way?
01:09:37 · Claire Vo
[01:09:37] S1: I think less about hardening memory and more about, one, managing context. So when I start to think, wow, we've been talking about this thing for a long time, doing a check-in and saying, "Make sure to write all this to your memory in case it gets compacted or Make sure our to do list is updated with the latest," is a really good use case. And there's some hooks that you can call inside of OpenClaw to do that automatically. And I think about it, maybe again, going back to managing an employee. If you leave a meeting, somebody takes the action items. So if you're leaving a discussion about a topic, make sure you check in and say, "Do you have all the action items? Are the notes written down in a place that we can work on?" Again, it's just like this operational hygiene that goes a long way to making your OpenClaw efficient. And then the thing that I find it forgets the most that people don't think about enough is what tools it can use and how. And so I've heard a million times, "I can't read that email address." And I'm like, "Dude, you can definitely read that email address." And so there is a tools.md that lists all the tools it has access to and how you want it to use the tools. While I don't recommend people edit the soul by hand, it's sometimes really useful to edit the tools document by hand because there's just some nuances in how you might want it to read your calendar or search the web, or I use Linear as tasking kind of substrate for my agents, how you might want it to use Linear. And so the tools markdown file is a really useful one.