Lenny's Podcast: Product | Career | Growth
01:21:47 · Lenny Rachitsky, Claire Vo
[01:21:47] S0: And the voice note there is like in Telegram, you could just say do voice note into an actual audio file. It's not like transcribed to text? [01:21:55] S1: No, you do just an audio. You hold the thing down, and then it does, I think it does whisper transcription and then reads it. And then it can talk back to you in a very robotic voice if you want. [01:22:05] S0: Any other tips? Anything else that you think might be helpful for folks? Any other challenges you ran into that might be helpful?
01:22:10 · Claire Vo
[01:22:10] S1: It's a little complicated to get things set up in a group chat. Again, I think the team has done a really good job over the last couple weeks of having it closed by default and then progressively opened as you understand more and can... It's like an escape room. If you can make your OpenClaw open, then you deserve its powers and can be trusted with its powers. And so setting up group chats can be a little bit of trial and error. You really have to unlock some settings, giving it access to tools, having it run code. All those things are pretty complicated. But my tip, and it's a little technical, but I promise you it's helpful, is install Claude Code or Codex on the same computer you're running your OpenClaw on and Claude Code the God mode administrator of your OpenClaws. So open up Claude Code, point it at the docs, say, "I have OpenClaw installed here, and Polly says she can't connect to email. Go fix." And Claude Code, because it's so good at writing code and OpenClaw is just mostly configuration code, can go in, read the docs and say, "Oh, you have this field here named ABC and it's supposed to be X, Y, Z. I've gone ahead and fixed it." Claude Code can also do that replicate agent work, the brain transplant job where you can say, "Hey, in OpenClaw, I have this agent Polly. I want to fracture off her memory and take all the family related stuff into Finn. Can you do that for me?" " And Claude pretty good at that. So you can use Claude Code as a surgeon and manager of your OpenClaw.
01:23:53 · Lenny Rachitsky, Claire Vo
[01:23:53] S0: That is so interesting. And the metaphor I was imagining is a brain surgeon going in and fixing things. So this is essentially like if OpenClaw just runs into some problem, you could have Claude look at it and figure things out. [01:23:54] S1: Yep. [01:24:06] S0: And to run Claude, it's the same situation. You install it using the terminal. There's a command you can find on Google, and then you run it in the OpenClaw directory. How do you find even where that lives? [01:24:16] S1: Yeah. So the OpenClaw directory is at your route. So basically when you open up terminal, it's going to drop you into your home. And in that home is OpenClaw. In that home is all your file system. So be careful. Yeah, I would just open up Claude Code there and say OpenClaw is @.openclaw. Oh, other magic trick that again people might not know is your OpenClaw files are in a hidden folder in Mac. And so I think you press command shift period when you're in Finder, and it'll show you your hidden files. And that's how you find find OpenClaw. [01:24:51] S0: Any other tips? Anything else along those lines?
01:24:53 · Claire Vo
[01:24:53] S1: It all sounds very technical. I really think read the docs is... The docs will save you. It's not that hard once you understand what's happening behind the scenes. So read the docs, keep your expectations tempered, narrow the scope of any of your agents. Definitely give your agents nice little avatars and color coding. I think that's really fun. And then the other thing is just, I say have a polite and positive relationship with your agents. One of the things that I think we can all get pulled into is when we have frustrating AI experiences, it can bring out the worst of us in some ways. I find myself being like, "Ah, literally I just told you this. " And I actually found myself the other day typing an angry message to one of my OpenClaws. I was like, "This is so frustrating. We've gone through this a million times. I don't know why this is a problem." And I was like, "That would not be effective on an employee." It would be a totally ineffective mechanism by which to manage an employee. Why would I think it would be an effective mechanism to manage an agent which is trained with a bunch of data from humans? And so I do think that, again, just bringing this manager's mindset to how you use these things, how you scope their roles, how you onboard them, how you onboard them technically, how you train them, how you give them more trust, I say bring those skills into it, again, not because we're going to personify the AI agents, but because I think that is why I have been so successful with these tools is because I have 20 years plus of management experience. I know how to make an employee successful. That is what you need to make these agents work. You don't need the technical skills. We can figure that out. You need role scoping or design, like voice. How do we talk to customers? How do we talk to each other? The rest of it's easy to follow. The rest of it, we can give to Claude Code to figure out.
01:26:58 · Lenny Rachitsky
[01:26:58] S0: The other interesting piece of that is just a lot. We just had this guest post by Molly Graham about the waterline model, and it was this pitch that most of the problems in your team, people jump to it's the person's fault, but most of it is like, "Okay, they don't actually understand their job. They don't understand what success looks like. They don't understand the goal." Or there's team overlap with who's responsible. So it's not like the person it's structural issues generally, and that feels like the same situation here. If your bot is doing the wrong thing, it's not that it's dumb, it just doesn't have the context, doesn't know what you want it to do.
01:27:32 · Claire Vo, Lenny Rachitsky
[01:27:32] S1: Yeah. And it's so funny because in an agentic system, that line is so clear. You can actually go into its file system and say, "Does it have this information? Should it be able to do this job?" And what you realize is no matter how you have communicated to it or explained to it that API, the Yapper API, while effective, is lossy. It's lossy. You tell somebody to do something, they forget it. You tell them to do it two weeks ago, they definitely forget it. And so making that effective communication onboarding so clear, so literal in the file system is very interesting because it is a reflection of how good am I at tasking my team? How good are my systems? How robust is my documentation? Because if this magical AI system with endless resources and infinite coding abilities can't figure out which projects are doing well and aren't, how am I supposed to figure that out? How is somebody I hire supposed to figure that out? So it is like almost maybe we need to do an RPG where you just go through the paces as a manager. And if you can make your agents effective, then you get promoted into human management. [01:28:50] S0: You mentioned Hilary Gridley, and she's big on this, just how much of the manager skillset translates to AI and agents.