Author: Baoyu
This is another 40-minute interview with Peter Steinberger, the author of ClawdBot/OpenClaw, hosted by Peter Yang.
Peter is the founder of PSPDFKit and has over 20 years of iOS development experience. After the company received a strategic investment of 100 million euros from Insight Partners in 2021, he chose to "retire." Now, his developed Clawdbot (now renamed OpenClaw) is a huge success. Clawbot is an AI assistant that can chat with you through WhatsApp, Telegram, and iMessage, connecting to various applications on your computer.
Peter describes Clawbot as:
“It’s like a friend living in your computer, a bit weird but smart enough to scare you.
In this interview, he shares many interesting viewpoints: why complex agent orchestration systems are "slop generators," why "running AI 24 hours" is a vanity metric, and why programming languages are no longer important.
One Hour Prototype, 300,000 Lines of Code
Peter Yang asks him what Clawbot is and why the logo is a lobster.
Peter Steinberger does not directly answer the lobster question but tells a story. After "retiring," he fully immersed himself in vibe coding—having AI agents help you write code. The problem is, the agent might run for half an hour or stop after two minutes to ask you a question. You go out for a meal and come back to find it has already gotten stuck, which is frustrating.
He wanted something that could check the status of his computer from his phone at any time. But he didn’t take action because he thought it was too obvious; big companies would definitely do it.
“By last November, when no one had done it, I thought, forget it, I’ll do it myself.
The initial version was extremely simple: connect WhatsApp to Claude Code. Send a message, and it calls the AI to send back the result. It was done in an hour.
Then it "came to life." Now Clawbot has about 300,000 lines of code and supports almost all mainstream messaging platforms.
“I think this is the direction of the future. Everyone will have a super powerful AI that walks with you through life.
He said, "Once you give AI access to your computer, it can basically do anything you can do."
That Morning in Morocco
Peter Yang says now you don’t have to sit in front of the computer staring at it; just give it commands.
Peter Steinberger nods but wants to talk about something else.
Once, he was in Morocco celebrating a friend's birthday and found himself using Clawbot the whole time. Asking for directions, finding restaurant recommendations—these were small matters. What really surprised him was that morning: someone tweeted that there was a bug in one of his open-source libraries.
“I just took a picture of the tweet and sent it to WhatsApp.
The AI understood the content of the tweet and recognized it as a bug report. It checked out the corresponding Git repository, fixed the issue, submitted the code, and then replied to that person on Twitter saying it was fixed.
“At that moment, I thought, can it really do that?
There was an even crazier instance. He was walking down the street, too lazy to type, so he sent a voice message. The problem was, he hadn’t even programmed voice message support for Clawbot.
“I saw it showing 'typing' and thought, this is it. But it replied to me normally.
Later, he asked the AI how it did that. The AI said: I received a file without an extension, so I looked at the file header and found it was in Ogg Opus format. You have ffmpeg on your computer, so I used it to convert it to WAV. Then I looked for whisper.cpp, but you didn’t have it installed, so I found your OpenAI API key and used curl to send the audio for transcription.
Peter Yang, after hearing this, said: These things are really capable, though a bit scary.
"Much stronger than the web version of ChatGPT; it’s like a ChatGPT unchained. Many people don’t realize that tools like Claude Code are not just good at programming; they are capable of handling any problem.
The Legion of Command Line Tools (CLI)
Peter Yang asks him how those automation tools were built, whether he wrote them himself or let AI write them.
Peter Steinberger laughs.
For the past few months, he has been expanding his "CLI legion." What are agents best at? Calling command line tools, because that’s all the training data consists of.
He built a CLI that accesses all Google services, including the Places API. He created one specifically for searching memes and GIFs, so the AI can send memes when replying to messages. He even made a tool to visualize sound, wanting the AI to "experience" music.
“I even hacked into the local food delivery platform's API, so now the AI can tell me how long until the food arrives. There’s also one that reversed the Eight Sleep API, allowing control of my bed's temperature.
[Note: Eight Sleep is a smart mattress that can adjust the surface temperature, and the official API is not open.]
Peter Yang follows up: Did you have AI help you build all these?
“The most interesting part is, I spent 20 years doing Apple ecosystem development at PSPDFKit, specializing in Swift and Objective-C. But after coming back, I decided to change tracks because I was tired of Apple controlling everything, and the audience for Mac apps is too narrow.
The problem is, switching from one tech stack you’re proficient in to another is painful. You know all the concepts but don’t know the syntax. What’s a prop? How do you split an array? Every little question requires research, and you feel like an idiot.
“Then AI came along, and all of that disappeared. Your system-level thinking, architectural ability, taste, and judgment of dependencies are what truly matter, and now they can easily transfer to any field.
He paused:
"Suddenly, I felt like I could build anything. Language doesn’t matter; what matters is my engineering mindset.
Controlling the Real World
Peter Steinberger begins to demonstrate his setup. The list of permissions he gives to the AI is astonishing:
Email, calendar, all files, Philips Hue lights, Sonos speakers. He can have the AI wake him up in the morning, gradually increasing the volume. The AI can also access his security cameras.
“One time, I had it watch for strangers. The next morning it told me: 'Peter, someone is here.' I looked at the footage, and it had been taking screenshots of my couch all night because the camera quality was poor, making the couch look like someone was sitting on it.
In his apartment in Vienna, the AI can also control the KNX smart home system.
“It can really lock me out of my house.
Peter Yang asks: How are these connected?
“Just talk to it. These things are very capable; it will find the API itself, it can Google, and it will look for keys in your system.
Users' applications are even crazier:
- Some have it shop at Tesco
- Some have it place orders on Amazon
- Some have it automatically reply to all messages
- Some have added it to family group chats as a "family member"
“I had it help me check in on the British Airways website. It was like a Turing test, operating a browser on an airline's website; you know how inhuman that interface is.
The first time it took almost 20 minutes because the whole system was still rough. The AI needed to find his passport in his Dropbox, extract the information, fill out the form, and go through human verification.
"Now it only takes a few minutes. It can click the 'I am human' verification button because it is controlling a real browser, and its behavior pattern is indistinguishable from a human.
80% of Apps Will Disappear
Peter Yang asks: For ordinary users who just downloaded it, what are some safe entry-level uses?
Peter Steinberger says everyone’s path is different. Some install it and immediately start using it to write iOS apps, while others go straight to managing Cloudflare. One user installed it for himself in the first week, for his family in the second week, and started creating an enterprise version for his company in the third week.
“I installed it for a non-technical friend, and he started sending me pull requests. He had never sent a pull request in his life.
But what he really wants to say is a bigger picture:
“If you think about it, this thing could replace 80% of the apps on your phone.
Why still use MyFitnessPal to track diet?
“I have an infinitely resourceful assistant that already knows I made a bad decision at KFC. I send a photo, and it saves it to the database, calculates the calories, and reminds me to go to the gym.
Why still use an app to set the temperature of Eight Sleep? The AI has API access and can adjust it directly. Why still use a to-do list app? The AI can remember for you. Why still use an app to check in for flights? The AI can do it for you. Why still use a shopping app? The AI can recommend, place orders, and track.
“There will be a whole layer of apps that will slowly disappear because if they have an API, they are just services that your AI will call.
He predicts that 2026 will be the year many people start exploring personal AI assistants, and big companies will also enter the market.
"Clawbot may not be the final winner, but this direction is correct.
Just Talk to It
The topic shifted to AI programming methodologies. Peter Yang mentioned that he wrote a popular article called "Just Talk to It" and wanted to hear Peter Steinberger elaborate on it.
Peter Steinberger's core point is: don't fall into the "agentic trap."
“I see too many people on Twitter discovering that agents are powerful, then wanting to make them even more powerful, and they fall down the rabbit hole. They build all sorts of complex tools to accelerate workflows, but in the end, they are just building tools, not creating anything truly valuable.
He has fallen into this trap himself. Early on, he spent two months building a VPN tunnel just to access a terminal on his phone. He did it so well that once, while having dinner with friends, he was vibe coding on his phone the whole time instead of participating in the conversation.
"I had to stop, mainly for my mental health.
Slop Town
What has been driving him crazy lately is a orchestration system called Gastown.
“An ultra-complex orchestrator running a dozen or twenty agents simultaneously, communicating and dividing tasks. There are watchers, overseers, mayors, and pcats (possibly referring to 'civilians' or 'pet cats' as filler roles), and I don’t even know what else.
Peter Yang: Wait, there’s a mayor?
“Yes, there’s a mayor in the Gastown project. I call this project 'Slop Town.'
There’s also the RALPH mode (a “use-and-throw” single-task loop mode, referring to giving the AI a small task, discarding all contextual memory after completion, resetting everything, and then entering a dead loop)…
“This is basically the ultimate token burner. You let it run all night, and the next morning you get ultimate slop.
The core issue is: these agents lack taste. They can be frighteningly smart in some ways, but if you don’t guide them and tell them what you want, the output is garbage.
“I don’t know how others work, but when I start a project, I only have a vague idea. Through the process of building, playing, and feeling, my vision gradually becomes clearer. I try some things, some don’t work, and then my ideas evolve into their final form. My next prompt depends on the current state of what I see, feel, and think.
If you try to write everything into the initial specifications, you miss out on this human-machine loop.
“I don’t know how you can create good things without feelings and taste involved.
Someone on Twitter boasted about a “fully RALPH-generated” note app. Peter replied, saying: Yes, it looks like it was generated by RALPH; no normal person would design it that way.
Peter Yang summarized: Many people run AI 24 hours not to create apps but to prove they can make AI run for 24 hours.
"It's like a size comparison contest without a reference point. I also let a loop run for 26 hours and felt proud at the time. But this is a vanity metric, completely meaningless. Just because you can build everything doesn’t mean you should build everything, nor does it mean it will be good.
Plan Mode is a Hack
Peter Yang asked him how he manages context. Does the AI get confused in long conversations, requiring manual compression or summarization?
Peter Steinberger said this is an “old mode problem.”
“Claude Code still has this issue, but Codex is much better. On paper, it might only have 30% more context, but it feels like 2-3 times more. I think it’s related to the internal thinking mechanism. Now, most of my feature development can be done within a single context window, with discussion and building happening simultaneously.
He doesn’t use worktrees because that’s “unnecessary complexity.” He simply checks out several repositories: clawbot-1, clawbot-2, clawbot-3, clawbot-4, clawbot-5. He uses whichever is free, completes testing, pushes to the main branch, and synchronizes.
“It’s a bit like a factory; if they’re all busy. But if you only open one, the wait time is too long, and you can’t get into a flow state.
Peter Yang said this is like a real-time strategy game; you have a team attacking, and you have to manage and monitor them.
Regarding plan mode, Peter Steinberger has a controversial viewpoint:
“Plan mode is a hack that Anthropic had to add because the model is too impulsive; it jumps straight into writing code. If you use the latest model, like GPT 5.2, you’re just having a conversation with it. 'I want to build this feature; it should be like this and that; I like this design style; give me a few options; let’s discuss first.' Then it will propose, you discuss, and reach a consensus before taking action.
He doesn’t type; he talks.
"I mostly talk to it.
Discord-Driven Development
Peter Yang asked him what the process is for developing new features. Do you explore the problem first? Do you make a plan first?
Peter Steinberger said he did something “possibly the craziest thing I’ve ever done”: he connected his Clawbot to a public Discord server, allowing everyone to chat with his private AI, complete with his private memories, in a public setting.
“This project is hard to describe in words. It’s a mix of Jarvis (the AI assistant from Iron Man) and the movie 'Her.' Everyone I demonstrate it to in person is super excited, but posting pictures with text on Twitter just doesn’t catch fire. So I thought, why not let people experience it themselves?
Users ask questions, report bugs, and make requests in Discord. His current development process is: take a screenshot of a Discord conversation, drag it into the terminal, and tell the AI, “Let’s talk about this.”
“I’m too lazy to type. If someone asks, 'Do you support this or that?' I let the AI read the code and write an FAQ.
He also wrote a crawler that scans the Discord help channel at least once a day, letting the AI summarize the biggest pain points, and then they fix them.
No MCP, No Complex Orchestration
Peter Yang asked: Do you use those fancy things? Multi-agent, complex skills, MCP (Model Context Protocol), and so on?
“My skills are mostly life skills: tracking diet, grocery shopping, that kind of stuff. Very little in programming because it’s not needed. I don’t use MCP or any of those things.
He doesn’t believe in complex orchestration systems.
“I’m in the loop; I can create products that feel better. Maybe there are faster methods, but I’m already close to the bottleneck not being AI; I’m mainly limited by my own thinking speed, occasionally by the time it takes to wait for Codex.
His former co-founder at PSPDFKit, a former lawyer, is now also sending him PRs (pull request).
“AI allows people without a technical background to build things, which is amazing. I know some people oppose this, saying the code isn’t perfect. But I treat pull requests as prompt requests; they convey intent. Most people don’t have the same system understanding and can’t guide the model to the optimal result. So I’d rather get the intent, do it myself, or rewrite based on their PR.
He marks them as co-authors but rarely directly merges others' code.
Find Your Own Path
Peter Yang summarized: So the core point is, don’t use slop generators; keep humans in the loop because human brains and taste are irreplaceable.
Peter Steinberger added:
“Or rather, find your own path. Many people ask me, 'How did you do it?' The answer is: you have to explore for yourself. Learning these things takes time and requires making your own mistakes. It’s like learning anything else, but this field changes particularly fast.
Clawdbot can be found at clawd.bot and on GitHub. Clad has a W, C-L-A-W-D-B-O-T, like a lobster claw.
(Note: ClawdBot has been renamed OpenClaw)
Peter Yang said he also has to give it a try. He doesn’t want to sit in front of the computer chatting with AI; he wants to be able to give it commands while out with the kids.
"I think you’ll like it," Peter Steinberger said.
Peter Steinberger's core points can be summarized in two sentences:
- AI is powerful enough to replace 80% of the apps on your phone.
- But without human taste and judgment in the loop, the output is garbage.
These two statements may seem contradictory, but they point to the same conclusion: AI is a lever, not a replacement. It amplifies what you already have: system thinking, architectural ability, and intuition for good products. If you don’t have these, running multiple agents in parallel for 24 hours is just mass-producing slop.
His practice itself is the best proof: a 20-year iOS veteran programmer built a 300,000-line project in TypeScript within a few months, relying not on learning the syntax of a new language but on those language-agnostic skills.
“Programming languages are no longer important; what matters is my engineering mindset.”
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
