Humans in on the Ralph Loop

🧔 Heads Up! This post was written with human assistance using Adam (me) and Claude Code .

Been stumbling into rad YouTube rabbit holes recently. Last night I was watching a YouTube video of Pookie setting up and using OpenClaw — I've been trying to figure out how useful OpenClaw is, or if the risk-to-value ratio is just too steep. There are so many attack vectors, and it seems like the more you mitigate them, the less powerful the tool is. I want to understand what's possible, before investing too much time. Pookie bought a Mac Mini, did the whole thing, and spent days just getting it to connect to his speakers and play music.

Days. To play music out of speakers.

The other thing I realized - after going through the install process myself, was ... it's not simple. And it's kinda buggie. To do it securely. To understand what I'm getting into. I have a hard time imagining non-technical people doing this (maybe a post or video on that soon). But, alas, I digress. I this post is about ... Ralph, or is it Geoffrey. Maybe both.

The Ralph Loop

As I was describing - Pookie had OpenClaw develop a DJ/Audio skill. But it had bugs, so he looked at the code, and it was .. garbage. He decided to refactor it using a pattern called the Ralph Loop, which has become popular lately. And when he mentioned it, he briefly flashed an image of a YouTube video by Geoffrey Huntley. Who is, I think, important. And, as it turns out, he's the inventor of the Ralph Loop (Or maybe it was his 9 year old 🤔).

The Ralph Loop, at its core, is dead simple. But I hadn't seen it done exactly the way Geoff does it (which is ironic, since he created it) with a one line bash command. It's a for loop:

for i in $(seq 1 $ITERATIONS); do claude -p "$PROMPT" --allowedTools "$TOOLS"; done

Or, from his site:

while :; do cat PROMPT.md | claude-code ; done

It's not a new concept, in-fact, Anthropic released a plugin to mimic the Ralph Loop in Claude Code, and it's just old patterns, with different actors. Instead of handing off the specs to my engineers (humans), I'm writing them collaboratively with my agents, then handing them off to other agents. So it's write requirements > specs, break them into detailed tasks > specs, then tell the agent to spawn an agent for each task, and update the main requirements doc, as it finishes. So each task remains small, preventing context rot, and the agents can run autonomously and report back for the other agents to note progress. There's a little more nuance, but that's basically it. And it can run, unattended, for .. hours, and in some cases days.

Geoffrey Huntley's YouTube video illustrating the Ralph Loop is a unique, strange, and compelling YouTube video. His website, and YouTube channel are gold.

Humans Should Not Be Writing Code Anymore

That shift happened quick. The last few months especially have been transformational, and I think most companies are going to take a while to catch up to this reality. Some SWE's probably too.

But humans should not be writing code anymore.

As an engineer, it can be hard to accept. But if you're a curious engineer, and love new tech, I think there's still opportunity. My job is totally different now. And I'm kind of enjoying it at the moment. That might change next week. In fact it probably will. But right now, I'm excited about what I can do with these tools, and energized to see what people like Geoffrey are doing with them.

Why Engineers Still Matter

But engineers still matter, or at least that's what Geoffrey Huntley posits - that engineering skill is still a requirement. And I tend to agree, at least for now.

I've been gobbling up his blog. Watching him work reminds me — or maybe illustrates is the better word — why engineers are still valuable. Claude turned soccer moms in to software developers. But if you give Codex/Claude Code to an engineer, the outcomes are next level. And if you want to build something beyond a lunch money tracker, you're still going to hire Geoffrey, not the vibe coder soccer mom.

For example Huntley, right now is working with his agents to write a new programming language and a compiler for it. He's not writing any of the code himself. He's using the Ralph Loop method to build it.

He supervises, and interjects when it takes a wrong direction. He calls this approach "human on the loop" instead of "human in the loop." It's a subtle but important distinction.

And when he wakes up to a broken codebase — and he does — he reviews it, can this be salvaged? Or do I git reset --hard, adjust the prompt, and kick off the loop again?

That's engineering. Just not the way it used to be.

The Rabbit Hole Is the Point

Last night I was looking for YouTube videos about OpenClaw, but I ended up discovering a workflow pattern and a creator whose thinking is reshaping how I approach my own work. One reference, the one flash of Geoffrey Huntley's video in the corner of the screen — that changed my week.

That's how it works. You follow the thread. You let one thing lead to another. And sometimes you find something that reframes everything.

Maybe this post does that for someone else. That would be cool.

---

Adam Daum is an agentic engineer and AI architect. He runs Weststack, LLC, an agentic AI and software engineering company, and writes about building practical AI solutions at adamdaum.com.