Conversation with ClawdBot Founder: AI is a Lever, Not a Replacement

Organized by: Baoyu

This is another 40-minute interview with Peter Steinberger, author of ClawdBot/OpenClaw, hosted by Peter Yang.

Peter is the founder of PSPDFKit, with nearly 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, the Clawdbot he developed (now renamed OpenClaw) is a huge hit. Clawbot is an AI assistant that can chat with you through WhatsApp, Telegram, and iMessage, connected to various applications on your computer.

Peter describes Clawbot like this:

"It's like a friend living in your computer, a bit weird, but frighteningly smart.

In this interview, he shares many interesting views: why complex agent orchestration systems are "slop generators," why "running AI for 24 hours" is a vanity metric, and why programming languages no longer matter.

One-Hour Prototype, 300,000 Lines of Code

Peter Yang asked him what Clawbot actually is and why the logo is a lobster.

Peter Steinberger didn't directly answer the lobster question but told a story. After "retiring," he returned and fully immersed himself in vibe coding—letting AI agents write code for you. The problem was, an agent might run for half an hour or stop after two minutes to ask a question. You come back from a meal to find it stuck long ago, which is annoying.

He wanted something to check his computer's status on his phone anytime. But he didn't build it because he thought it was too obvious, and big companies would surely do it.

"By November last year, no one had done it, so I thought, forget it, I'll do it myself.

The initial version was extremely simple: connect WhatsApp to Claude Code. Send a message, it calls the AI, and sends the result back. It took an hour to set up.

Then it "came to life." Now Clawbot has about 300,000 lines of code and supports almost all major messaging platforms.

"I think this is the future direction. Everyone will have a super-powerful AI that follows them throughout their 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 said that now you don't need to sit in front of the computer staring at it; just give it instructions.

Peter Steinberger nodded, but he wanted to talk about something else.

Once, while celebrating a friend's birthday in Morocco, he found himself using Clawbot constantly. Asking for directions, restaurant recommendations—these were small things. What really surprised him was that morning: someone posted a tweet on Twitter saying there was a bug in one of his open-source libraries.

"I just took a photo of the tweet and sent it to WhatsApp.

The AI read the tweet, understood it was a bug report. It checked out the corresponding Git repository, fixed the issue, committed the code, and then replied to that person on Twitter saying it was fixed.

"I thought, is this even possible?

Another time was even more incredible. He was walking on the street, too lazy to type, so he sent a voice message. The problem was, he never built voice message support for Clawbot.

"I saw it display 'typing,' and thought, oh no. But then it replied normally.

He later asked the AI how it did it. The AI said: I received a file without an extension, so I looked at the file header and found it was Ogg Opus format. You have ffmpeg on your computer, so I used it to convert to WAV. Then I looked for whisper.cpp, but you didn't have it installed, but I found your OpenAI API key, so I used curl to send the audio for transcription.

Peter Yang听完说: These things are really resourceful, though a bit scary.

"Much stronger than the web version of ChatGPT. This is like ChatGPT unleashed. Many people don't realize that tools like Claude Code aren't just good at programming; they are resourceful with any problem.

Command-Line Tool (CLI) Army

Peter Yang asked him how those automation tools were built, whether he wrote them himself or had the AI write them.

Peter Steinberger laughed.

He had been expanding his "CLI army" for months. What are agents best at? Calling command-line tools, because the training data is full of that.

He built a CLI to access the entire Google services, including the Places API. Built one specifically for finding memes and GIFs, so the AI could send memes when replying to messages. He even made a tool to visualize sound, wanting the AI to "experience" music.

"I also hacked into the API of a local food delivery platform; now the AI can tell me how long until the food arrives. Another one reverse-engineered the Eight Sleep API to control the temperature of my bed.

[Note: Eight Sleep is a smart mattress that can adjust bed surface temperature; the official API is not open.]

Peter Yang pressed: Did you have the AI help you build all these?

"The most interesting thing is, I did Apple ecosystem development at PSPDFKit for 20 years, Swift, Objective-C, very specialized. But after coming back, I decided to switch tracks because I was tired of Apple controlling everything, and making Mac apps has too narrow an audience.

The problem was, switching from one精通的技术栈 to another is painful. You understand all the concepts but don't know the syntax. What's a prop? How do you split an array? You have to look up every little problem; you feel like an idiot.

"Then with AI, all that disappeared. Your systems thinking, architectural skills, taste, judgment about dependencies—these are the truly valuable things, and now they can easily migrate to any field.

He paused:

"Suddenly I felt like I could build anything. The language doesn't matter anymore; what matters is my engineering thinking.

Controlling the Real World

Peter Steinberger began demonstrating his setup. The list of permissions he gives the AI is staggering:

Email, calendar, all files, Philips Hue lights, Sonos speakers. He can have the AI wake him up in the morning, slowly increasing the volume. The AI can also access his security cameras.

"Once I told it to watch for strangers. The next morning it told me: 'Peter, someone is here.' I looked at the录像, it had been screenshotting my sofa all night because the camera quality was poor, and the sofa looked like a person sitting there.

In his Vienna apartment, the AI can also control the KNX smart home system.

"It could really lock me out.

Peter Yang asked: How are these connected?

"Just by talking to it. These things are resourceful; it will find the API itself, Google things, look for keys in your system.

Users'玩法 are even crazier:

• Someone had it shop online at Tesco
• Someone had it place orders on Amazon
• Someone had it automatically reply to all messages
• Someone added it to a family group chat as a "family member"

"I had it check me in on the British Airways website. This is like a Turing test; operating a browser on an airline website, you know how anti-human that interface is.

The first time it took almost 20 minutes because the整套系统 was still rough. The AI needed to find his passport in his Dropbox, extract information, fill out the form, pass the CAPTCHA.

"Now it only takes a few minutes. It can click the 'I'm human' verification button because it's controlling a real browser, its behavior pattern is no different from a human's.

80% of Apps Will Disappear

Peter Yang asked: For a regular user who just downloaded it, what are some safe ways to get started?

Peter Steinberger said everyone's path is different. Some people install it and start using it to write iOS apps immediately; others go straight to managing Cloudflare. One user installed it for themselves the first week, for their family the second week, and started working on an enterprise version for their company the third week.

"After I installed it for a non-technical friend, he started sending me pull requests. He had never sent a pull request in his life.

But what he really wanted to talk about was the bigger picture:

"If you think about it, this thing might replace 80% of the apps on your phone.

Why still use MyFitnessPal to record diet?

"I have an infinitely resourceful assistant; it already knows I made a wrong decision at KFC. I send a photo, it saves it to the database, calculates calories, reminds me I should go to the gym.

Why still use an app to set the Eight Sleep temperature? The AI has API access, it can adjust it for you directly. Why still use a to-do app? The AI remembers for you. Why still use an app to check in for a flight? The AI does it for you. Why still use shopping apps? The AI can recommend, order, track.

"A whole layer of apps will slowly disappear because if they have an API, they are just services your AI will call.

He predicted that 2026 will be the year many people start exploring personal AI assistants, and big companies will also enter the field.

"Clawbot might not be the final winner, but this direction is correct.

Just Talk to It

The topic turned to AI programming methodology. Peter Yang said he wrote a very popular article called "Just Talk to It" and wanted to hear him elaborate.

Peter Steinberger's core point is: Don't fall into the "agentic trap."

"I see too many people on Twitter discovering agents are powerful, then wanting to make them more powerful, and falling down the rabbit hole. They build各种复杂的工具 to accelerate workflows, but end up just building tools, not building truly valuable things.

He fell into it himself. Early on, he spent two months building a VPN tunnel just to access the terminal on his phone. He did it too well; once he was having dinner with a friend at a restaurant, he spent the whole time vibe coding on his phone instead of participating in the conversation.

"I had to stop, mainly for mental health.

Slop Town

What recently drove him crazy was an orchestration system called Gastown.

"A super complex orchestrator, running a dozen or twenty agents simultaneously, they talk to each other, divide labor. There are watchers, overseers, mayors, pcats (probably meaning 'civilians' or 'pet cats' or other filler roles), I don't even know what else.

Peter Yang: Wait, there's a mayor?

"Yes, the Gastown project has a mayor. I call this project 'Slop Town.'

And the RALPH pattern (a "use and discard" single-task loop pattern, meaning giving the AI a small task, discarding all context memory after completion, resetting everything to zero, and then an infinite loop)...

"This is the ultimate token burner. You let it run all night, and the next morning you get the ultimate slop.

The core of the problem is: These agents don't have taste yet. They are frighteningly smart in some aspects, but if you don't guide them, don't tell them what you want, what comes out is garbage.

"I don't know how others work, but when I start a project, I only have a vague idea. During the process of building, playing, feeling, my vision gradually becomes clear. I try some things, some don't work, and then my idea evolves into its final form. My next prompt depends on the current state I see, feel, and think about.

If you try to write everything into前期规格说明, you miss out on this human-in-the-loop cycle.

"I don't know how you can make good things without feeling, without taste being involved.

Someone on Twitter was炫耀 a "fully RALPH-generated" notes app. Peter replied: Yes, it looks like it was generated by RALPH; no normal person would design it like this.

Peter Yang summarized: Many people run AI for 24 hours not to make an app, but to prove they can make AI run for 24 hours.

"It's like a measuring contest without a reference point. I've also let loops run for 26 hours and felt proud at the time. But this is a vanity metric, meaningless. Being able to 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. When the conversation gets long, the AI gets confused; does it need manual compression or summarization?

Peter Steinberger said this is a problem of the "old mode."

"Claude Code still has this problem, but Codex is much better. On paper, it might only be 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 completed within one context window; discussion and building happen simultaneously.

He doesn't use worktrees because that's "unnecessary complexity." He simply checks out several copies of the repository: clawbot-1, clawbot-2, clawbot-3, clawbot-4, clawbot-5. Use whichever is idle, test when done, push to main, sync.

"It's a bit like a factory, if they are all busy. But if you only have one, the waiting time is too long, you can't get into a flow state.

Peter Yang said this is like a real-time strategy game; you have a team attacking, you have to manage and monitor them.

About plan mode, Peter Steinberger has a controversial view:

"Plan mode is a hack Anthropic had to add because the model is too impulsive and rushes to write code immediately. If you use the latest models, like GPT 5.2, you just talk to 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 chat first.' Then it will propose, you discuss, reach a consensus, and then start.

He doesn't type; he talks.

"I talk to it most of the time.

Discord-Driven Development

Peter Yang asked him what his process is for developing new features. Explore the problem first? Make a plan first?

Peter Steinberger said he did something "probably the craziest thing I've ever done": he connected his own Clawbot to a public Discord server, letting everyone talk to his private AI, with his private memories, in public.

"This project is hard to describe in words. It's like a mix of Jarvis and the movie 'Her.' Everyone I演示 in person gets super excited, but posting pictures with text on Twitter just doesn't take off. So I thought, just let people experience it themselves.

Users ask questions, report bugs, and request features 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. Someone asks 'do you support this or that,' I have the AI read the code and then write an FAQ.

He also wrote a crawler that scans the Discord help channel at least once a day, has 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), etc.?

"Most of my skills are life skills: recording diet, grocery shopping, that kind of stuff. Very few programming ones, not needed. I don't use MCP, don't use any of those things.

He doesn't believe in complex orchestration systems.

"I'm in the loop; I can make products that feel better. Maybe there are faster methods, but I'm already fast to the point where the bottleneck isn't the AI; I'm mainly limited by my own thinking speed, occasionally by the time waiting for Codex.

His former PSPDFKit co-founder, a former lawyer, is now also sending him PRs (pull requests).

"AI allows people without technical backgrounds to build things; it's magical. I know some people oppose it, saying this code isn't perfect. But I treat pull requests as prompt requests; they convey intent. Most people don't have the same system understanding to guide the model to the optimal result. So I prefer to get the intent and do it myself, or rewrite based on their PR.

He marks them as co-authors but rarely directly merges others' code.

Find Your Own Way

Peter Yang summarized: So the core takeaway is, don't use slop generators, keep the human in the loop, because the human brain and taste are irreplaceable.

Peter Steinberger added:

"Or rather, find your own way. Many people ask me 'how do you do it,' the answer is: you have to explore it yourself. Learning these things takes time, requires making your own mistakes. It's like learning anything, except this field changes特别快.

Clawdbot can be found at clawd.bot and on GitHub. Clad with a W, C-L-A-W-D-B-O-T, like lobster claws.

(Note: ClawdBot has been renamed OpenClaw)

Peter Yang said he also has to try it. Doesn't want to sit in front of the computer chatting with AI; wants to give it instructions anytime while outside with the kids.

"I think you'll like it." Peter Steinberger said.

Peter Steinberger's core views can be summarized in two sentences:

1. AI is already powerful enough to replace 80% of the apps on your phone
2. But without human taste and judgment in the loop, the output is garbage

These two sentences seem contradictory but actually point to the same conclusion: AI is a lever, not a replacement. It amplifies what you already have: systems thinking, architectural skills, intuition for good products. If you don't have these, running再多智能体并行 for 24 hours is just mass-producing slop.

His practice itself is the best proof: a 20-year veteran iOS programmer built a 300,000-line code project in TypeScript within a few months, relying not on learning the syntax of a new language, but on those language-agnostic things.

"Programming languages don't matter anymore; what matters is my engineering thinking."