<p>Given OpenClaw’s hype and incredible pace, one cannot help but wonder how Peter, the creator, is able to achieve such a phenomenal impact all by himself. I read through his tweets, videos, blog posts, and everything that he shared publicly and here’s his playbook.</p>
<h2>Who is Peter Steinberger?</h2>
<p>Before OpenClaw, Peter founded PSPDFKit, a PDF SDK used by companies like Dropbox and Autodesk. He sold the company for nine figures after running it for over 13 years. Then he retired for three years.</p>
<p>In April 2025, he came back. AI coding had gotten good enough to reignite his spark. He built 43 projects before OpenClaw went viral. The project started as a simple WhatsApp to Claude Code relay that took him just one hour to build. Today, it’s 300,000 lines and used by so many people to automate their day to day.</p>
<h2>The Philosophy</h2>
<h3>Close the Loop</h3>
<p>Peter believes code is the perfect domain for AI because it’s verifiable. The agent writes code, runs it, sees errors, fixes them, and repeats. No human needed in the debug cycle. This is why AI excels at coding compared to writing or design where verification is subjective.</p>
<h3>You Are the Taste Layer</h3>
<p>Agents are what Peter calls “spiky smart.” They’re brilliant at execution but terrible at vision. They don’t know what matters. Your job is navigation and curation.</p>
<blockquote>
<p>“What agents don’t really do yet is have taste. If you don’t navigate them well, if you don’t have a vision, it’s still going to be slop.”</p>
</blockquote>
<blockquote>
<p>“My next prompt depends on what I see and feel and think about the current state of the project.”</p>
</blockquote>
<h3>The Agentic Trap</h3>
<p>Here’s where many developers go wrong. Agents are so capable that people fall into building ever-more-sophisticated AI tooling instead of actual products.</p>
<p>Peter calls out specific anti-patterns:</p>
<ul>
<li><strong>GasTown</strong>: An orchestrator with “watchers, overseers, and a mayor.” He calls it “slop town.”</li>
<li><strong>RALF</strong>: One task, loop, trash context, repeat. “The ultimate token burn machine.”</li>
</ul>
<blockquote>
<p>“Just because you can build everything doesn’t mean you should or that it’s going to be good.”</p>
</blockquote>
<p>When someone tweeted about a “fully RALF’d” Mac app, Peter replied: “Yeah, it looks RALF.” You could visually tell no sane human designed it.</p>
<h3>The Real Limiting Factor</h3>
<blockquote>
<p>“I’m mostly limited by thinking about it, sometimes a little bit by waiting for Codex. But mostly it’s my thinking that’s the limiting factor.”</p>
</blockquote>
<h2>The Workflow</h2>
<h3>Parallel Execution</h3>
<p>Peter runs 5-10 agents simultaneously. Each lives in a separate git checkout: <code>clawdbot-1</code>, <code>clawdbot-2</code>, <code>clawdbot-3</code>, and so on. His screen is split with multiple terminals. Each checkout is either free or working on something. When work is done, he pushes to main.</p>
<blockquote>
<p>“Feels like a factory if they’re all busy.”</p>
</blockquote>
<blockquote>
<p>“Like StarCraft - main base and side bases giving you resources.”</p>
</blockquote>
<p>He switches between agents while waiting, never blocking on any single task. When an agent finishes, he reviews the output, gives the next instruction or merges.</p>
<h3>The Spec-Driven Development Method</h3>
<p>This is Peter’s core workflow for larger features:</p>
<p><strong>Step 1: Context Dump</strong>
Dump the entire codebase into Gemini with its 1M token context window. Gemini handles massive context ingestion better than anything else.</p>
<p><strong>Step 2: Architectural Discussion</strong>
Have a conversation about what you want to build. Discuss tradeoffs, approaches, edge cases. Let Gemini help think through the architecture.</p>
<p><strong>Step 3: Generate Spec</strong>
The output is a ~500-line specification document. This is the complete blueprint.</p>
<p><strong>Step 4: Hand to Codex</strong>
Feed the spec to Claude Code with a simple instruction: <code>build spec.md</code>. The agent implements while you move to the next task.</p>
<h3>Full Gate: Local Validation</h3>
<p>Before any merge:</p>
<pre><code>lint → build → tests → <span class="hljs-built_in">push</span>
</code></pre>
<p>This happens locally, not in CI. Don’t wait for CI to tell you something is broken. The agent should run the full validation loop before considering work done. Close the loop.</p>
<h2>Tool Philosophy</h2>
<h3>CLIs Over MCPs</h3>
<blockquote>
<p>“I don’t use MCPs or any of that crap.”</p>
</blockquote>
<p>Why CLIs win:</p>
<ul>
<li>Can chain: <code>tool1 | tool2 | tool3</code></li>
<li>Can filter: <code>tool | jq '.field'</code></li>
<li>Composable with standard Unix patterns</li>
<li>Agents trained extensively on CLI usage</li>
</ul>
<p>MCPs are “silly.” You can’t pipe output, can’t filter or transform, and you’re locked into specific integrations. Extra complexity for no benefit.</p>
<blockquote>
<p>“What are agents good at? Calling CLIs. That’s what they trained for.”</p>
</blockquote>
<h3>CLIs Peter Built</h3>
<p>He built command-line tools for everything:</p>
<ul>
<li>All of Google Workspace plus Places API</li>
<li>Meme and GIF lookup for message replies</li>
<li>Sound visualization (“I wanted it to experience music”)</li>
<li>Food delivery status and ETA</li>
<li>Eight Sleep bed temperature control</li>
<li>Home automation: Philips Hue, Sonos, cameras, door locks</li>
<li>Flight check-in via browser automation</li>
</ul>
<h2>Model Selection</h2>
<h3>GPT-5.x / Codex: The Primary Choice</h3>
<blockquote>
<p>“Takes 10x longer but almost always gets it right.”</p>
</blockquote>
<p>The advantages are significant:</p>
<ul>
<li>Better context retention (feels like 2-3x more despite being 30% more on paper)</li>
<li>No need for plan mode - just have a conversation</li>
<li>Won’t build until you say “build”</li>
<li>Discussion and building happen in the same session</li>
<li>Most features now fit in one context window</li>
</ul>
<h3>Claude Code: Secondary</h3>
<p>Faster but more “trigger-happy.” Needs plan mode to prevent premature building. Context fills faster. Still useful for certain tasks.</p>
<h3>Gemini: Spec Generation</h3>
<p>Best for massive context dumps with its 1M token window. Use for initial architecture discussion. Generates the spec that other models implement.</p>
<h3>The Plan Mode Reality</h3>
<blockquote>
<p>“Plan mode was a hack that Anthropic had to add because the model is so trigger-happy and would just run off and build code.”</p>
</blockquote>
<p>With Codex, you just have a conversation. Say “let’s discuss” or “give me options.” The model naturally waits for your “build” signal. Peter talks using voice. He doesn’t type.</p>
<h2>Prompting and Communication</h2>
<h3>PRs Are Prompt Requests</h3>
<p>Peter’s ex-lawyer business partner now sends him pull requests. Peter treats them as intent signals, not implementations.</p>
<blockquote>
<p>“I treat pull requests more as prompt requests because they convey the idea… most people don’t have the same system understanding, so they’re not able to derive the model to give optimal results. I just extract the intent and rebuild it myself.”</p>
</blockquote>
<p>He reads prompts more than he reads code.</p>
<h3>Voice Over Typing</h3>
<p>Peter uses voice messages extensively. There was a pivotal moment in Morocco when he sent a voice note without having built voice support. The agent figured it out on its own:</p>
<blockquote>
<p>“I saw a file with no extension. Looked at the header - it was OGG audio. Found FFmpeg on your computer, converted to WAV. Looked for whisper.cpp but you didn’t have it. Found your OpenAI key, used curl to hit their API, got transcript, replied to you.”</p>
</blockquote>
<blockquote>
<p>“It’s like unshackled ChatGPT.”</p>
</blockquote>
<h2>Feedback Loops</h2>
<h3>Discord-Driven Development</h3>
<p>Peter scrapes the OpenClaw Discord help section daily. He asks the model: “What are biggest pain points?” He screenshots Discord conversations and drags them directly into his terminal.</p>
<blockquote>
<p>“Can you read the code and write a new FAQ entry?”</p>
</blockquote>
<p>Features start from observing where users have friction.</p>
<h3>Ship Fast, Learn Faster</h3>
<p>He built 43 projects before one went viral. Each project builds intuition. Failure is the teacher.</p>
<blockquote>
<p>“You have to make your own mistakes.”</p>
</blockquote>
<blockquote>
<p>“Prompting is a skill like programming.” You get better by doing, not reading.</p>
</blockquote>
<h2>Key Takeaways</h2>
<ol>
<li>
<p><strong>Close the loop.</strong> Agents must verify their own work. Code is perfect for AI because it’s testable.</p>
</li>
<li>
<p><strong>You provide the taste.</strong> Without human vision and curation, output is slop.</p>
</li>
<li>
<p><strong>Avoid the agentic trap.</strong> Build products, not infinitely sophisticated AI tooling.</p>
</li>
<li>
<p><strong>Run agents in parallel.</strong> Multiple checkouts, multiple terminals. Like StarCraft bases.</p>
</li>
<li>
<p><strong>Use CLIs, not MCPs.</strong> Agents are trained on command-line tools. Let them compose.</p>
</li>
<li>
<p><strong>Talk, don’t type.</strong> Voice input, natural conversation. No fancy orchestration.</p>
</li>
<li>
<p><strong>Ship constantly.</strong> 43 projects before viral success. Each failure teaches.</p>
</li>
<li>
<p><strong>Your thinking is the bottleneck.</strong> Not the AI.</p>
</li>
</ol>
<hr>
<h2>Sources</h2>
<p><strong>Twitter</strong>: <a href="https://twitter.com/steipete" target="_blank" rel="noopener nofollow">@steipete</a></p>
<p><strong>Blog</strong>: <a href="https://steipete.me/posts/" target="_blank" rel="noopener nofollow">steipete.me</a></p>
<p><strong>Podcasts</strong>:</p>
<ul>
<li><a href="https://youtu.be/8lF7HmQ_RgY" target="_blank" rel="noopener nofollow">Pragmatic Engineer Interview</a> - Deep dive on coding workflow and model selection</li>
<li><a href="https://youtu.be/AcwK1Uuwc0U" target="_blank" rel="noopener nofollow">Peter Yang Interview</a> - Focus on OpenClaw, personal AI assistant, and the agentic trap</li>
</ul>


Everything about vibe coding 🤖
Tools, tips, tricks, workflows, best practices, you name it.
Let the vibes guide you

Vibe Coding

Peter Steinberger, creator of OpenClaw and former PSPDFKit founder, shares his approach to building at exceptional speed using AI coding agents. His methodology centers on running 5-10 agents in parallel across separate git checkouts, using Gemini for spec generation and Claude Codex for implementation. He advocates for CLI tools over MCPs for composability, emphasizes closing verification loops so agents can self-correct, and positions humans as the "taste layer" providing vision while agents handle execution. Key insights include avoiding over-engineered AI tooling, using voice input over typing, validating locally before merging, and treating user feedback from Discord as direct feature requests. He built 43 projects before achieving viral success, demonstrating that shipping constantly and learning from failures is essential.

How Peter Steinberger Ships at 10x Speed

<p>“spiky smart” is such a clever term to describe an agents behaviour.</p>


<p>It’s easy to be 10x when you do not care too much about code quality and you’re not working on a big company repo that already has coding standards, guidelines for everything, frameworks, etc…</p>
<p>I did both: let the AI go wild and do whatever it thinks is the best, create files automatically, update them etc… In that case - yes, your output is in fact 10x’d, I’d argue maybe even more than 10x which is mindblowing!</p>
<p>But the moment you have to pay attention to what it outputs, and follow the above mentioned standards, it’s more around 2x. That’s still amazing, 2x is a massive improvement, but I reject around 80% of Claude Code recommendation when I’m not “vibe” coding.</p>


<p>That starcraft reference resonates with me.</p>
