A few weeks ago I came across interviews with Peter Steinberger — the guy behind OpenClaw, the fastest-growing open source project on GitHub. He shipped 8,471 commits across 48 repos in 72 days. That's about 118 commits per day, running 5–10 AI coding agents simultaneously.
I've been on a similar tear, just with more constraints. I left Atlassian in early 2025, moved my family from San Francisco to Zurich, and once we settled in I went heads-down on my own projects. Since November 19th I've made 4,092 contributions across 18 repos in 101 days — about 40 per day, with a 99% active-day rate. My peak day was 135 contributions. I'm doing this while raising a 6-year-old and a 1-year-old.
This is about what I learned from watching him work publicly, and how it changed my own approach.
Take Codex seriously
This is the biggest one. I was fully Claude-pilled. I liked the conversational feel — Claude has a very user-friendly default personality, it suggests next steps, it feels like working with someone rather than issuing commands. I hadn't bothered to look beyond it.
Steinberger kept saying: Codex just works. Claude jumps in and starts doing stuff and says it's done without even reading the codebase. Codex is more thorough. It reads the code first.
I've since started using Codex regularly through Yep Anywhere (my mobile agent supervisor) and he's right. Codex is better at:
- Fixing stuck tests — Claude will sometimes spin in loops on a failing test. Codex tends to actually figure out what's wrong.
- Backend tasks — it's just better at this kind of work, independent of the sandbox.
- Thoroughness — it reads more of the codebase before acting. Claude's eagerness is a feature until it isn't.
Honestly, I started using Codex because I hit my $200/month Claude token limit and was forced to try something else.
Codex out of the box is very dry — it does the task and stops. The single most useful customization I made was telling it to suggest something logical to do after completing a task. That immediately made it feel way more usable. Claude does this by default, which is probably why I stuck with it for so long.
Now I pick the model based on the task. Claude for conversational, exploratory work where I'm thinking out loud. Codex for backend, tests, and anything where I want it to actually read the codebase first.
Prompt requests, not pull requests
This was the insight that stuck with me the most. Steinberger's take: don't send me a pull request, send me a prompt request.
When someone sends a PR, you can usually see immediately that you would never do this yourself. They're changing stuff they shouldn't, adding weird markdown files, making choices that don't fit the project's direction. It's not that you'd spend a long time reviewing — it's that the PR is obviously wrong because the contributor doesn't know the constraints that live in your head, the things that aren't encoded in the codebase.
But a prompt? You can run it yourself, with your context. The code is reproducible. The judgment about what the code should do isn't.
The phone as a forcing function
Steinberger popularized the "control your agents from Telegram" workflow. I built a whole product around the same idea — Yep Anywhere lets me supervise Claude Code and Codex sessions from my phone.
I never would have prioritized mobile access if I could just sit at my desk all day. Desktop is more productive simply because you can have more things on screen at once. But on your phone, you can't get lazy and skip ahead to manually test — so you invest in the agent's ability to verify its own work. Tests become the priority because they're the only option.
Meta-tooling compounds
Every time Steinberger hit a limitation with AI agents, he built a tool to fix it. Agents can't see the screen? Build Peekaboo. Agents stuck in loops? Build Oracle. Need to message from agents? Build CLIs for iMessage, WhatsApp, Gmail.
This compounds. Every tool you build makes the next task easier, which means you build more things, which means more tools. It's a flywheel.
My workflow
Here's what a typical session looks like:
- Assess the task, pick the right model. Claude Code for interactive work, Codex for backend or when Claude is getting stuck.
- Start the session, frame the problem. Watch the first few moves to make sure it's on the right track.
- Once I trust the trajectory, go YOLO mode — let the agent execute without approval on every step.
- Context-switch — start another session on a different project, or go play with the kids.
- Come back, review the diffs, commit the work.
I'll run 2–5 sessions concurrently. I'm always on the main branch — no worktrees, no feature branches. If two tasks might touch the same code, I just sequence them. Sessions are short — a prompt, 10–20 minutes of execution, sometimes up to an hour for long test runs.
The numbers in context
| Me | Steinberger | |
|---|---|---|
| Period | 101 days | 72 days |
| Total contributions | 4,092 | 8,471 |
| Per day | 40 | 118 |
| Concurrent sessions | 2–5 | 5–10 |
| Supervision style | Hands-on | More autonomous |
| Shipping to | App stores, npm | Open source |
The point of the table isn't to make it a race. It's that 118 commits/day sounds superhuman until you see the inputs — more concurrent agents, more hours, no kids. The workflow itself isn't magic. It's accessible to anyone willing to learn it.
Who can do this
I think the reason there aren't more people doing this is that it requires an unusual intersection:
- Senior engineering skills — you need enough experience to architect systems and review AI output critically
- Financial runway — this doesn't pay (yet) while you're building
- Free time — or at least, fragmented time you can fill with short agent sessions
- Genuine obsession — I basically do nothing else besides code and spend time with my family
Most developers have one or two of these, not all four.
What I'm building
For context, here's what those 4,092 contributions went toward:
- JSTorrent — cross-platform BitTorrent client (Chrome extension, Tauri desktop, Android, CLI). Rebuilding my old Chrome App that historically was a top paid app in the Chrome Web Store.
- Yep Anywhere — mobile supervisor for AI coding agents. What I use daily to manage all of this from my phone.
- Tilefun — Roblox/Minecraft-style game engine I'm building with my daughter.
- claw-starter — personal assistant framework with browser, email, and messaging skills.
- Plus a handful of game prototypes and experiments.
If you're interested in the mobile agent workflow, try Yep Anywhere: npm i -g yepanywhere. And go watch Steinberger's interviews — the ideas are genuinely good, and he's generous about sharing them.