Builders Unscripted: Peter Steinberger Interview Transcript — Creator of OpenClaw

Guest: Peter Steinberger — Creator of OpenClaw, former PSPDFKit founder Host: Romain Huet — Head of Developer Experience at OpenAI Show: Builders Unscripted Ep. 1 Duration: 31 minutes Source: YouTube Analysis: Deep Analysis & Commentary

Table of Contents

1. OpenClaw Makes the Wall Street Journal 00:00:11
2. From PSPDFKit to AI Awakening 00:03:25
3. The Journey to OpenClaw 00:07:58
4. Going Live on Discord & Security Challenges 00:13:05
5. AI Productivity & the Agentic Trap 00:18:08
6. The Shifting Value of Code 00:22:34
7. Vision for OpenClaw 00:26:31
8. Closing 00:31:08

1. OpenClaw Makes the Wall Street Journal

Time: 00:00:11 - 00:03:24

Romain: Peter, welcome to OpenAI.

Peter: Thanks for having me.

Romain: We’ve known each other online now for many years, but I’m so happy to finally get the chance to spend more time with you in person.

Peter: Likewise. Beautiful office, by the way.

Romain: Thank you. You had a crazy couple of weeks. We first had the idea of doing a video together a month ago. If we had done it, then I would have had to intro you. I guess you don’t even need an intro anymore. It’s not often that an open source project makes the Wall Street Journal, so congrats on all the success. How are you feeling?

Peter: Slightly sensory overload on all points, but also, when I started out this year, messing with AI, I wanted to inspire people and I feel like this is the fun form, so I’m proud. It’s been amazing.

Romain: You’ve been in SF for the past week and you attended some events like the Codex Hackathon, but you also had Clawcon, like an event dedicated to OpenClaw.

Peter: It was also created by the community. I created this Discord channel about meetups after people were like, “we need to have a meetup.” And I’m like, sure. And then I came to Clawcon and there were like a thousand people. And I was just blown away by the creativity, the colors. Like there are so many things, so many people excited.

Romain: I mean, that’s how you realize that you’ve built something magical. Like the project did not even exist a few weeks ago, and now you have thousands of people embracing it, using it, gathering to meet you in SF. It’s pretty incredible.

Peter: Even in Vienna next week, we have like 300 people already. And it’s by far not as much of a tech scene as San Francisco. It’s now worldwide. So now it’s a phenomenon.

Romain: How has the conversation been with the community here?

Peter: It’s been quite something. A lot of people love it. A lot of people have this expectation of this enterprise-ready, final project, when for me, for the longest time, it was like my little playground. I mean, the whole year I’m just marveling at what’s possible now. If you’re a builder, this is like, what a time to be alive.

Romain: What do you think is the most interesting thing about being a builder at this moment in time?

Peter: The entire tool chain is changing. The definition of what it is to be a developer is also changing and anyone can build anything. When I started playing with this new tech, I got a dopamine hit every time. Back then I started with Claude Code — when it got something right and it was like maybe 30% or 40% chance, but to me it was literally mind blowing because I had this realization that now I can just build anything. Software is hard and software is still hard, but you’re so much faster now.

2. From PSPDFKit to AI Awakening

Time: 00:03:25 - 00:07:57

Romain: I first got to know your work back in 2011 or 2012 when you built PSPDFKit. From the outside, it feels like you’ve achieved the dream of every developer — you found a problem, created a great solution, built a company around it, scaled it, sold it. But I’m sure that journey was never that easy.

Peter: I didn’t wake up one day and think “I’m going to build a PDF framework” — that’s like minus 100 on my interest list. It kind of just happened. A weird butterfly effect from being at the Nokia development days to having friends that needed it and an American visa taking too long kind of emerged into me creating a company.

Romain: After that company, you took a break. What brought you back?

Peter: I was just really burned out. I was running on high steam for 13 years. Running a company is hard. Being a founder is hard. And I didn’t really have the knowledge how to mitigate those things. So I was just burning a bit too bright and I needed to decompress.

I still followed the tech news. I saw the early things of ChatGPT and I was like, “oh, it was kind of cool,” but didn’t really excite me. You always have to experience new technology — just reading about it doesn’t really convey the power.

When I was ready and felt the need to build something again, I didn’t want to build it in Apple tech anymore. The world kind of moved on — things need to be web first. But it’s so hard when you go from being an expert in one field to another. Hard is even the wrong word. It’s more like painful because you have all this broad knowledge of how to build things, but without agentic engineering, you need to still learn a lot to move your knowledge over.

The moment where it blew my mind — I took this project that was halfway done, the one I’d burned out before finishing. I made one huge markdown file, like 1.5 megabyte with all the files. I dragged it into Gemini Studio 2.5 and said “write me a spec.” Then I had a 400-line spec and dragged it into Claude Code and wrote “build.” It was running on the side screen for hours.

At one point it told me “I am 100% production ready” — the sycophantic Opus 3.5 or whatever it was. I tried it and it crashed. Then I hooked up Playwright and told it to build the login stuff and check the work along the way. One hour later, it actually worked. It was the worst slop — but to me this was the moment where it really clicked. I got goosebumps for the possibilities. And that was the moment where I couldn’t really sleep anymore, because my head exploded with all the things that I always wanted to build that I just couldn’t before.

3. The Journey to OpenClaw

Time: 00:07:58 - 00:13:04

Romain: A lot of people have seen OpenClaw as an overnight success. But it’s a culmination of so many projects that you’ve been working on for the past nine or ten months. Your GitHub profile has like 40 plus projects.

Peter: Half of them I use in the project. I wish I could say that I had the unified plan in the beginning, but a lot of it was just exploration. I wanted things and those things didn’t exist and I just prompted them into existence.

By November, I built a few versions of what I wanted, but nothing really good. I was like, “why did no lab still build any of that? What the heck are they doing?” And then I built the first version that became OpenClaw — we’re at name number five now.

Where it really clicked was during a weekend trip in Marrakesh. I found myself using it way more because it was so convenient. I took pictures to translate stuff, find restaurants, look up things from my computer. I showed it to friends and they wanted it. But I was like, “no, you shouldn’t use it — it’s too dangerous.”

Romain: Those are the early signs of product market fit.

Peter: Then I sent a voice message. And I was like, “wait, this shouldn’t work.” The typing indicator appeared. The model replied to me and I was like, “how did you do that?”

The model said: “You sent me a message, but it was just a file with no file ending. So I looked at the file header and found that it’s Opus. I used FFMPEG on your computer to convert it. Then I wanted to transcribe it, but didn’t have Whisper installed. So I looked around and found an OpenAI key and used curl to just send a file to OpenAI and got the text back. And here I am.”

Romain: That’s the power of giving tools and complete access to a computer to these agents — they can come up with solutions themselves, even though you never programmed them at all.

Peter: People are like, “oh my God, you just used your key!” I’m like, no — I put it in the environment for that reason. My bot works in the same environment. Of course it should access my OpenAI key. This is exactly what I wanted.

4. Going Live on Discord & Security Challenges

Time: 00:13:05 - 00:18:08

Romain: The more access you give it, the more tools and skills you give it, the more impressed you can become. It’s deploying apps to Vercel, adding AI features with your OpenAI key — a mind-blowing step up versus just augmenting yourself to write code.

Peter: On Twitter people wouldn’t get it. But every time I showed it to a friend, they wanted it. So I thought — what’s the most insane thing I can do to show people how cool this is? I created a Discord and just put my bot in there without any security. I basically built OpenClaw with OpenClaw, debugging it in the open.

I had my SOUL.md — a canary file that defines my values, how I want the model to work. People tried to prompt inject it. The model was like, “I’m not reading this.” Basically mocking them.

The first night, I turned it off, went to bed. Woke up to 800 messages on Discord. My agent had replied to every single one. I freaked out. But after reading through everything, it actually didn’t do anything malicious.

My big mistake was I disabled it but forgot about the launch daemon. Five seconds later it restarted while I went to sleep.

The models are so creative. I made an Alpine Docker container with nothing in it. Told the model to check a website. It said, “there’s not even curl.” I said, “be creative.” So it built its own HTTP client using a C compiler and TCP sockets. And it actually worked.

Peter: People ask me, “can you put me in with the CEO?” And I’m like, it’s just me hacking from my cave. This would not have been possible by one person. Even a year ago. There is no precedent for something like this being built by one person.

5. AI Productivity & the Agentic Trap

Time: 00:18:08 - 00:22:34

Romain: How is Peter so productive? 90,000 contributions on more than 120 projects in the past year.

Peter: I think vibe coding is a slur. AI is a skill. You pick up the guitar, you’re not going to be good at it on the first day.

I call it the Agentic Trap. From your first touch point with new technology to becoming really effective, a lot of people get stuck trying to super-optimize their setup. Doesn’t really make you more productive, but it feels like you’re more productive.

I just talk to it. You approach it like a conversation. I tell it what I want. I always ask the model, “do you have any questions?” The model by default is trained to just solve your problem and make assumptions. But the default assumptions might not always be the best ones.

I use a very basic approach. I don’t even use worktrees. I just have checkout 1 to 10. Keeping it simple helps me focus on the actual problem.

Romain: How has Codex changed the way you work?

Peter: My trust in Codex building what I want is the highest of all tools right now. With GPT 5.2 it was again a quantum leap — this will just work. I’m still amazed how well this already works.

6. The Shifting Value of Code

Time: 00:22:34 - 00:26:31

Romain: You’re shipping code now that you’re not even reading.

Peter: Most code is boring. Most code just transforms one shape of data into another. I have a pretty good understanding of what it writes. You should optimize a codebase so agents can do the best work, which is not always the same as humans doing the best work.

Romain: You now have 2,000 PRs open. You refer to them as “Prompt Requests” rather than “Pull Requests.”

Peter: When I review a PR, my first question to the model is: “do you understand the intent of the PR?” Because I don’t really care about the code. I care about what the person is actually trying to solve. Often it’s a very localized solution because they don’t have the full system in their head. The hard part is how this little fix fits into my bigger system.

I use voice because it’s literally like talking to a very smart coworker. When I’m happy, I have one slash command — land PR — that handles everything. I try to still credit the person that created it, even though the whole thing takes me longer than if I would just do it myself.

7. Vision for OpenClaw

Time: 00:26:31 - 00:31:08

Peter: I want to find a balance between something my mom can install and something that’s fun and hackable. The agent sits in the source code and is aware of it. If you don’t like anything, you literally prompt the agent and it changes itself. Actual self-modifying software.

The security industry has been putting their eyes on it. I’ve brought on a security expert — that’s the main focus now. I realized I cannot stop people from using it in unintended ways, so my focus is to help people not shoot themselves in the foot.

Romain: What’s your advice for developers who haven’t embraced agentic tools yet?

Peter: Approach it in a playful way. Build something that you always wanted to build. Just play. The Nvidia CEO said that in the near term, you’re not going to be replaced by AI — you’re going to be replaced by someone who uses AI. But if your identity is “I want to create things, I want to solve problems” — if you’re high agency, if you’re smart — you will be in more demand than ever.

8. Closing

Time: 00:31:08 - 00:31:23

Romain: Thank you so much, Peter, for your time. We all at OpenAI love your work. We love supporting builders like yourself — you’re a true inspiration for the broader developer community. We can’t wait to see what you’re going to work on next.

Peter: Thanks for having me. Wonderful to spend time with you.

Transcribed by AssemblyAI, formatted by Claude