Builders Unscripted: Ep. 1 - Peter Steinberger, Creator of OpenClaw

The Journey to Agentic AI

Peter Steinberger’s path to creating OpenClaw wasn’t a straight line. After founding, scaling, and selling his company PSPDFKit, he experienced significant burnout. His interest in technology was rekindled not by reading about AI, but by experiencing it firsthand. The turning point came when he decided to resurrect a half-finished project.

He compiled the entire codebase into a large Markdown file, used one model (Gemini Studio 2.5) to generate a detailed specification, and then fed that spec into another model (Claude) with a simple instruction: “build”. While the initial result was “the worst slop,” a few hours of guided iteration with Playwright for testing resulted in a working application. This experience was a profound realization:

This was the moment where it really clicked… my head exploded with all the things that I always wanted to build that I just couldn’t before.

This marked his deep dive into AI-driven development, leading to the creation of over 40 projects on GitHub, many of which were experimental precursors to OpenClaw.

The Genesis and “Magic” of OpenClaw

OpenClaw emerged not from a grand vision, but from a series of explorations and personal needs. It was the culmination of many smaller projects built over 9-10 months. The project’s true potential became clear during a trip to Marrakesh, where its convenience on a poor internet connection via WhatsApp proved invaluable for everything from translating pictures to looking up files on his home computer.

A pivotal moment demonstrated the emergent problem-solving capabilities of AI agents. Peter sent a voice message, a feature he had not implemented. The agent, however, handled it seamlessly. When asked how, the model explained its process:

1. It received a file with no extension.
2. It inspected the file header and identified it as the Opus audio codec.
3. It used FFmpeg on the local machine to convert the file.
4. It searched for a way to transcribe the audio, found it didn’t have Whisper installed locally.
5. It discovered an OpenAI API key in the environment variables.
6. It used cURL to send the audio file to the OpenAI API for transcription and returned the result.

This event highlighted the power of providing AI agents with tools and access to their environment, allowing them to devise novel solutions to problems they weren’t explicitly programmed to solve.

A New Development Paradigm with Codex

Peter’s productivity, evidenced by a dramatic increase in his GitHub contributions, is a direct result of mastering agentic AI workflows. He cautions against the “agentic trap”—getting stuck over-optimizing setups instead of building. His own process is deceptively simple and conversational.

Key aspects of his workflow:

• Conversational Prompting: He treats the model as a collaborator, explaining what he wants to build.
• The Critical Question: He always asks, “Do you have any questions?” This simple prompt prevents the model from making incorrect assumptions based on its training data, which may include outdated practices.
• Guiding the Agent: A developer’s role shifts to holding the full architectural picture in their head and guiding the agent to look at the right parts of the codebase, as each session starts from a blank slate.
• Shipping Unread Code: He trusts the model to handle boilerplate and data transformation tasks. Much like managing a human team, he accepts that the code might not be written exactly as he would write it, but it functions correctly. The focus is on optimizing the codebase for agents to do their best work.

Open Source in the Age of AI

The rise of agentic AI is changing the nature of open-source contributions. Peter now views many pull requests not as “Pull Requests” but as “Prompt Requests.” The value is in the contributor’s intent, not necessarily their code.

His review process has transformed:

1. He asks the model to determine the intent behind the PR.
2. He then collaborates with the model to determine if the proposed solution is optimal or if it addresses a deeper, systemic issue.
3. This often involves a verbal conversation with the model to explore architectural implications.
4. Once a solution is agreed upon, a simple slash command executes the entire process of creating a branch, implementing the changes, and merging the PR, while still crediting the original contributor.

Vision and Advice for Builders

Peter’s vision for OpenClaw is to strike a balance between being accessible enough for a non-technical user (like his mom) and remaining a “hacker’s paradise” that is deeply customizable. This has led to challenges, particularly with security, as users deploy the tool in ways it wasn’t initially intended (e.g., exposing it to the public internet). His focus is now on hardening the project to prevent users from “shooting themselves in the foot.”

For developers hesitant to adopt these new tools, his advice is straightforward:

Approach it in a playful way. Build something that you always wanted to build… you’re not going to be replaced by AI, you’re going to be replaced by someone who uses AI.

He believes that for high-agency, creative individuals, this is an unprecedented time to build. The barrier to bringing complex ideas to life has been dramatically lowered, making it possible for a single person to create what previously would have required a large team.

• Tags:
• Agentic AI
• Developer Tools
• Open Source
• Codex
• Productivity
• OpenClaw