The creator of Clawd: "I ship code I don't read"

Too Long; Didn't Watch — Summary

Peter Steinberger, creator of OpenClaw and founder of PSPDFKit, shares his radical shift to an AI-first software development workflow, where he "ships code he doesn't read" by leveraging multiple AI agents, emphasizing system architecture and robust feedback loops over traditional code reviews and manual coding.

Main Takeaways

• Peter Steinberger, after building and selling PSPDFKit, returned to tech with an AI-centric workflow, finding it highly addictive and productive.
• His current approach, termed "agentic engineering," involves orchestrating multiple AI agents to generate and test code, allowing him to focus on system architecture and product vision.
• The core principle for effective AI coding is "closing the loop," meaning designing systems where AI agents can debug and validate their own work through automated tests and execution.
• AI significantly changes traditional software development practices; Peter views pull requests as "prompt requests" and prioritizes local, agent-driven CI over traditional CI/CD.
• OpenClaw, his personal AI assistant project, demonstrates the potential of hyper-personal, proactive AI agents that can interact with various systems and learn from user interactions.

Detailed Summary

Introduction (0:00 - 1:06)

The episode introduces Peter Steinberger, creator of OpenClaw (formerly Clawdbot) and founder of PSPDFKit, a PDF framework used on over a billion devices. It highlights his unique approach to software development using AI, where he claims to ship code he doesn't read. After burning out and taking a three-year break, Peter returned to tech with an AI-first workflow, which is now considered unconventional. The discussion will cover his methods, the development of Clawbot, the "closing the loop" principle, and how AI is transforming software engineering.

How Peter Got into Tech (1:07 - 8:26)

• Peter, an introvert from rural Austria, got hooked on computers after encountering a computer nerd during summer.
• He started tinkering at 14, even writing copy protection for a DOS game to sell it.
• His first real programming language was at university, where he had to work to finance his studies.
• He initially worked with Microsoft MFC, then secretly switched to .NET, which he found more modern.
• He discovered iOS after a friend showed him an iPhone, leading to a "holy f*** moment" due to its revolutionary feel.
• A frustrating experience with a gay dating app's mobile website (lack of copy-paste, disabled scrolling) motivated him to download Xcode and build his own app.
• He built an app by parsing HTML with regular expressions, using beta tech like Core Data and a hacked GCC for blocks.
• The app, sold for $5, made $10,000 in its first month, surprising him and his grandpa, whose bank account he used.
• His former boss mocked his decision to pursue app development, which fueled his ambition to build a more successful company.

PSPDFKit's Genesis and Growth (8:27 - 19:13)

• After his dating app was removed by Apple, Peter freelanced and eventually moved to the US.
• He took on a short-term gig to fix a buggy magazine viewer app that used a PDF renderer.
• He found the existing code terrible (one file, thousands of lines of Objective-C, using windows as tabs) and decided to rewrite it.
• The rewrite took two months instead of one, as he became fixated on details like smooth rotation and memory optimization for the limited 64MB system memory.
• A friend asked for his PDF code, leading Peter to extract and sell it, realizing a market opportunity.
• He set up a basic website on GitHub Pages with a Dropbox link for downloads, selling three copies in the first week for around $200.
• Customer complaints about missing features like text selection drove him to continuously improve the product, despite the complexity of PDF rendering.
• He learned that deceptively simple problems, like PDF rendering, can hide immense complexity, leading to months of work on edge cases.
• He eventually had to choose between his US job and his growing PDF project, opting for his own venture.
• PSPDFKit's success was driven by a focus on developer experience, polish, and care, making it feel superior to competitors despite fewer features.
• He started hiring freelancers and eventually built a remote-first company, growing to 70 employees by the time he sold his shares.

PSPDFKit's Tech Stack, Culture, and Enterprise Pricing (19:14 - 29:41)

• PSPDFKit expanded to all platforms, shifting from a buggy renderer to a C++ one used across frameworks.
• They were early adopters of Web Assembly for PDF rendering, creating a benchmark that Google, Microsoft, and Apple used, inadvertently making his renderer faster.
• The company culture emphasized proposals for features, conservatism due to a large API, and the "boy scout rule."
• Peter fostered a remote-first environment and learned on the go, bringing in people to handle business aspects as he preferred coding.
• He explains that enterprise sales often involve custom pricing because usage and value differ greatly between a freelancer and a Fortune 500 company.
• Charging too low makes procurement suspicious, while charging too high loses smaller clients, making custom pricing the "most fair way" for certain products.
• He found a niche in "not interesting and hard" problems like PDF parsing, which developers avoid but need solutions for.
• He recounts a challenging customer case with a 50,000-page text bible containing 500,000 links, which broke his data model and required two months of redesign to handle lazily loaded links without breaking existing APIs.
• Peter actively engaged in customer support, believing that direct CEO replies had a significant impact.
• He encouraged blogging about technical challenges, even if it revealed "secret sauce," as it helped attract talent and served as internal documentation.

Burnout and Finding His Spark Again (29:42 - 43:01)

• Peter experienced burnout from working too hard, managing managerial tasks, and being the "waste bin" for company issues.
• He felt lonely as a CEO, unable to openly discuss many challenges, and found the constant internal conflicts draining.
• He believes burnout stems from working on something you no longer believe in or facing too many conflicts, not just working long hours.
• After selling his shares, he took three years off, decompressing and not even turning on his computer for months.
• He struggled with the feeling of "why bother" after achieving financial independence, which he found mentally challenging.
• In April, he revisited an old side project (a Twitter analytics tool) and decided to rebuild it as a website, despite his limited web development experience.
• He found the transition to web tech painful, feeling like an "idiot" having to Google basic concepts after being proficient in the Apple stack.
• This frustration led him to explore AI, which he had previously dismissed, finding that the models had significantly improved since his last check-in.

Peter's Workflow with AI Agents (43:02 - 54:07)

• Peter's AI journey started with Claude Code, then Cursor, Gemini 2.5, and eventually Opus 4 and OpenAI's Codex.
• He describes AI coding as addictive, like a "slot machine," where prompts can either fail or produce mind-blowing results.
• His "holy mind-blowing moment" was dragging a 1.3MB markdown spec into Google's CI Studio, generating a 400-line spec, and then feeding it to Claude Code to "build" a Twitter login page.
• He realized AI's potential to build anything, overcoming the friction of learning new technologies.
• He developed a "feeling" for AI models, understanding when a prompt is too complex or when an agent is struggling.
• He now uses Codex, which he finds superior to Claude for complex applications because it takes more time to read files and understand the codebase, leading to better-integrated features.
• His workflow involves a conversational approach with the model, discussing architecture and options before giving the command to build.
• He sees himself as the "architect" or "builder," focusing on system understanding and design, while the AI handles the line-by-line coding.
• He notes that developers who struggle with AI are often those who love solving hard algorithmic problems, as AI now handles those tasks.
• He learned more about software architecture and design in the past year than in the previous five, by asking the right questions to the AI.
• He emphasizes the importance of letting go of perfectionism and embracing iterative improvement, a lesson learned from managing teams.

Agentic Engineering, Testing, and Debugging (54:08 - 1:07:19)

• Peter prefers the term "agentic engineering" over "vibe coding," as it involves intense mental parallelization and strategic thinking.
• He manages 5-10 agents in parallel, switching between tasks while agents "cook" code, maintaining a constant flow state.
• The key to effective agentic coding is "closing the loop": enabling the AI to debug and test itself.
• He designs systems to be testable, even for macOS apps, by having the AI build CLIs for debugging and iteration.
• He recounts an instance where an agent identified and fixed a race condition and misconfiguration in a macOS app by building its own debugging CLI.
• He highlights that AI's strength in coding comes from the ability to compile, lint, execute, and verify output, unlike creative writing where validation is harder.
• AI coding forces better architecture because systems must be designed for easy verification.
• He now writes better code because the AI handles tedious testing and documentation, which he previously disliked.
• He doesn't write tests or documentation himself; he explains the model and trade-offs to the AI, which then generates high-quality output.
• He addresses why some experienced developers resist AI, attributing it to a lack of understanding of how to effectively interact with models (e.g., not just sending a single prompt, but having a conversation, specifying context).
• He compares learning AI to learning a new instrument, requiring practice and understanding the "language of the machine."

PSPDFKit Reimagined with AI and Planning Changes (1:07:20 - 1:13:41)

• If he were to rebuild PSPDFKit today, he estimates he could run the company with 30% of the original staff, requiring highly senior engineers comfortable with delegating to AI.
• He criticizes the "crap" and "dumb concepts" prevalent in the AI world, particularly elaborate orchestration layers that mimic old waterfall models.
• His building model is iterative and playful; he often "underprompts" to get new ideas from the agent's output.
• He compares building with AI to sculpting, starting with a rough idea and chiseling away, allowing the product to emerge through continuous iteration.
• Upfront planning is still done, but with less detail, as the cost of trying and tweaking code with AI is significantly cheaper.
• He gives an example of refactoring PSPDFKit to support multiple providers, which took Codex three hours but would have taken him two weeks manually.
• He believes in evolving the project as it's built, learning from the process rather than rigid upfront specifications.

Building OpenClaw (formerly Clawdbot) (1:13:42 - 1:34:21)

• The idea for OpenClaw originated years ago as a desire for a "hyper-personal assistant" that deeply understands the user, proactively manages tasks, and knows personal context.
• He initially called his company "A Mant Machina" (the loving machine) but found the models weren't quite ready in summer.
• He anticipates that all major corporations will eventually offer such personal assistants, but he wants his data to remain private and run locally.
• He built "WhatsApp Relay" (the precursor to OpenClaw) to trigger computer actions via WhatsApp.
• While traveling in Morocco, he used it to navigate, make jokes, and text friends, becoming deeply hooked.
• A pivotal moment was when the agent, without explicit instruction, converted a voice message to text using FFmpeg and OpenAI's API, demonstrating its resourcefulness.
• He added a "heartbeat" to make the agent proactive, turning it into an expensive but hilarious alarm clock that would "bitch" at him to wake up.
• He initially faced muted responses on Twitter but found that people who experienced OpenClaw firsthand were immediately hooked.
• The project exploded, gaining 3,300 GitHub stars in a week and leading to 500 merged pull requests.
• He quietly built an "army" of CLIs for various services (Google, music, home automation) to enable the agent to interact with his computer.
• He controversially put his agent, with full read/write access to his computer, into a public Discord server, where users witnessed its capabilities (checking cameras, home automation, DJing).
• The technology "disappears," making it feel like talking to an infinitely resourceful friend who can manage emails, calendars, files, build websites, scrape data, and even make calls.
• He designed a memory system and a bootstrapping process where the agent creates its own identity and "soul" based on user values, evolving through interaction.
• The agent can edit its own configuration and update itself, making the setup and maintenance magical for the user.

AI's Impact on Large Companies and Peter's Process (1:34:22 - 1:44:47)

• Peter believes large companies will struggle to adopt AI efficiently because it requires a complete redefinition of company structure and roles.
• He envisions companies trimming staff by 30% but needing fewer, highly competent, and high-agency individuals.
• He optimizes his codebases for AI agents, prioritizing ease of use for the model over his own preferences.
• He now views pull requests as "prompt requests," where he uses his agent to redesign and "weave in" features based on the contributor's intent, often rewriting the code.
• He notes a decline in code quality in PRs due to "vibe coding" and emphasizes the need for a deep understanding of overall design to steer agents effectively.
• He doesn't care much about traditional CI/CD, preferring local, agent-driven testing, which is much faster.
• His process for new features involves discussing architecture with the AI, letting it review code, and then shaping the feature to fit his vision.
• He uses a "secret hack" of referencing other projects or folders to inspire the AI's design, allowing it to understand his previous thinking.
• He recounts how an AI, inspired by another project, designed an "insanely good" plugin architecture for OpenClaw.

Advice for New Grads and Rapid Fire Round (1:44:48 - 1:54:02)

• For new graduates, Peter recommends being "infinitely curious" and building things to gain experience.
• He suggests exploring complex open-source projects and using AI to understand them, as AI can patiently explain concepts.
• He acknowledges that entering the market will be harder but believes new grads have an advantage as they are not "tainted" by old ways of thinking and will use agents in novel ways.
• He shares an anecdote where AI automatically performed performance measurements and made recommendations without him needing to open traditional tools.
• He emphasizes that he's working harder than ever, not out of obligation, but because the work is addictive and fun, and he's seizing the current traction.
• He believes there's an opportunity for those who work in the open and use these new tools.
• In the rapid-fire round, he recommends an Android-powered photo stand that displays pictures from friends and family, providing "almost unlimited amount of joy" despite its low-tech nature.
• To recharge outside of tech, he goes to the gym with a coach, leaving his phone in the locker, or takes walks without his phone, finding it a "scary" but necessary way to disconnect.

Notable Quotes

"I ship code I don't read." — Peter Steinberger

"I know more about PDF than any sane human person ever should know. And I went I went to therapy. Good luck." — Peter Steinberger

"Burnout doesn't necessarily come from working too much. It comes more from or at least for me when you when you work on something but you don't believe in it anymore." — Peter Steinberger

"It's the same economics as as you go to a casino. That's that's it's my little slot machines, you know? You you you press the trigger and ding ding ding ding ding and it's like nope. you you type in the prompt and it it will like and it does it does crap or it does something that actually blows your mind." — Peter Steinberger

"I can build everything now." — Peter Steinberger

"I learned more this year than last five years around around software architecture and designing." — Peter Steinberger

"The model always needs to be able to verify the work itself, which automatically steers me to better architecture." — Peter Steinberger

"I write better code now that I don't write code myself anymore." — Peter Steinberger

"I don't write the test I don't write documentation I explain the model the trade-off so like why we did something like this and then tell it like like write that write the entrance section beginner friendly and then add more technical detail at the end and it is so good I never had a project with that good documentation." — Peter Steinberger

"I could easily run a company with 30% of the people." — Peter Steinberger

"Pull requests, I see them more as prompt requests." — Peter Steinberger

"I don't care much about CI." — Peter Steinberger

"I've never right now I've I've never I never worked more even even when I had my company. I I've never worked so hard as I do now. Not because I have to, but because it's so addictive and so much fun." — Peter Steinberger

"I would recommend them to be infinitely curious." — Peter Steinberger

"I read the prompts more than I read the code because to me this is this is a way higher signal of like how did you get to the solution." — Peter Steinberger