10 Principles of AI Coding from the OpenClaw Creator

I was genuinely struck by an interview with Peter Steinberger, the creator of OpenClaw (formerly Clawdbot) - the project that hit the fastest star growth in GitHub history.

Peter is a veteran who ran a 60-70 person company for 13 years, sold it, took a three-year break, and came back. His perspective on development in the AI era is fundamentally different from what most of us assume.

How It Started

His beginning was simple. There was no grand business plan. He just wanted to “play around with AI,” and ended up building a tool because he wanted to chat with his home computer via WhatsApp while he was out.

The Aha Moment

The defining moment came during a trip. He sent a voice message to his agent - but he had never coded voice support. The agent figured out the Opus file format on its own, found ffmpeg to convert it, located an API key, transcribed and translated the message, and sent back a reply. That was the moment he realized agents are “clever, resourceful beasts.”

Based on experiences like this, Peter distilled his 10 principles for AI coding.

Drop Perfectionism to Work with AI

Managing a team of 70 taught him to accept work that doesn’t match his personal style. Code that doesn’t align 100% with your taste but works correctly is good enough. That flexibility is now his biggest asset when collaborating with agents.

Design Systems Where Agents Verify Their Own Work

Peter calls this “closing the loop.” Compile, lint, run, verify - the agent handles all of it. When humans have to confirm intermediate steps, they become a bottleneck that slows everything down.

Pull Requests Are Dead - Welcome to Prompt Requests

The code itself matters less than the prompt that produced it. Peter rejects most external PRs, extracting only the core idea and recycling it as a prompt. His brother works the same way - a sign this pattern is already spreading.

Replace Code Reviews with Architecture Discussions

Even on Discord, Peter’s core team doesn’t talk about code. They discuss system structure, major decisions, and direction. The entire team has internalized the view that implementation details belong to the agent.

Run 5 to 10 Agents Simultaneously

Instead of grinding on a single task, queue multiple tasks in parallel. Plan, hand off to an agent, and immediately move to the next one. This is how Peter maintains a flow state throughout the day.

Spend a Surprising Amount of Time Planning

Peter ping-pongs extensively with agents during the planning phase - challenging, revising, pushing back, iterating until the plan feels solid. He prefers Codex over Claude Code for execution because Claude Code asks clarifying questions mid-run, which breaks the flow. When the plan is airtight, execution requires almost zero intervention.

Give Deliberately Vague Instructions

Over-specifying instructions constrains the AI to operate only within that scope. Leaving room intentionally lets the agent discover directions you hadn’t considered. In my own experience, this genuinely works - unexpected solutions surface regularly. That said, it’s not always the right approach.

Test Locally Instead of Waiting 10 Minutes for Remote CI

Waiting 10 minutes for a remote CI pipeline is wasted time. Design the system so agents run tests locally. The shorter the feedback loop, the faster you iterate.

Most Code Is Just Boring Data Transformation

The bulk of application code is moving data from one shape to another. There’s no reason to obsess over it - delegate it to the agent. Save your energy for system design, not data plumbing.

People Who Love Shipping Adapt Faster to AI

Developers who enjoy solving algorithmic puzzles actually struggle more with the AI transition. Those who care about outcomes and shipping products adapt quickly. I see this pattern consistently in people around me.

Peter’s View of the Future

He predicts that countless apps will disappear, leaving only APIs. Instead of opening MyFitnessPal to manually log food, you’ll send a photo to your agent, which will calculate calories and adjust your health goals automatically.

The Takeaway

There’s room for debate, but Peter’s 10 principles converge on a single direction: drop perfectionism, discuss architecture instead of reviewing code, let agents verify their own work, and run multiple agents in parallel.

All of this points toward building an environment where the developer doesn’t have to write the code themselves. If that’s where we’re headed, then the real skill in the AI era isn’t writing great code - it’s designing the system that solves problems without you writing any code at all.

References:

• Interview with Peter Steinberger (Part 1)
• Interview with Peter Steinberger (Part 2)