Avatar
AI
Craft
Systems

It is not a snake eating its tail

6 min read

TL;DR

  • An eighteen month old video that already reads like an archaeological find, and the three file workflow inside it.
  • Why the output of one AI being the input to the next is the whole trick.
  • The ouroboros is the wrong picture, because that snake gets smaller.
  • Newton stood on the shoulders of giants. This thing stands on its own.
  • Why it could not have built this without me, and the one job that never hands off.

The moment it broke open for me was watching Ryan Carson, a five time founder, explain how he actually works. It was May 2025, on Claire Vo's How I AI. He was not announcing anything. He was just showing his screen.

A 3-step AI coding workflow for solo foundersHow I AI, with Claire VoWatch the episode

Go and watch it now and it feels like handling something dug out of the ground. It is eighteen months old. The tool he is working in has changed beyond recognition, and he is careful with the model in a way almost nobody bothers to be careful any more. In this field, eighteen months is a geological era.

And yet nearly every serious way of working with AI I have seen this year is still that video's idea wearing better clothes. That is the thing about relics. The casing dates quickly. What is inside it does not.

What he showed was almost embarrassingly simple, which is why it landed.

Three files

He had three instruction files sitting in his project.

The first one tells the AI how to write a specification. He opens a fresh chat, says in one or two sentences what he wants to build, and points it at that file. Out comes a proper written spec, in language plain enough for a junior developer to follow. He reads it. He argues with it. He gets it right before a single line of code exists.

The second file turns that spec into a task list, small enough to finish one at a time. The third works the list, stopping after each task so he can say yes or no.

Spec, tasks, execution. Three steps, and he gave the prompts away for free. They are still on GitHub as ai-dev-tasks.

The output of one is the input to the next

In step one, the AI is not writing code. It is writing a document. And that document's only purpose is to be fed to the AI in step two. Which also writes no code. It writes a task list, whose only purpose is to be fed to the AI in step three.

Two thirds of that workflow is a machine writing instructions for a machine.

I had been treating AI as something I aim at a problem. Carson was treating it as something that could aim itself, as long as a person stood at each joint and said yes. Everywhere a human used to write the brief, the machine could write the brief and the human could approve it, and approving is faster than authoring by a distance.

It is not literally three different AIs. It is the same tool three times, with a clean context and a different instruction file each time, which is the whole point rather than a technicality. The expensive thing was never the model. It was the handover, and he had made the handover a file.

The wrong picture

Everybody reaches for the same image when they see this. The ouroboros, the snake with its own tail in its mouth. AI using AI, the circle closing, the thing consuming itself.

A snake with its jaws clamped on its own tail, the tail crushed where the fangs grip it, its body thinning all the way round, and nothing inside the circle.

It is the wrong picture, and the wrongness matters. A snake eating its tail is a closed loop that produces nothing. Whatever is inside that circle is all there is ever going to be, and it is getting smaller with every bite.

This is the opposite. Nothing is consumed. Each turn leaves something behind that was not there before. The spec did not exist, and now it does. The task list did not exist, and now it does.

Part of why people reach for the wrong picture is that nature barely offers a right one. Living things reproduce, and the offspring starts small and makes the whole climb again from nothing. A coral reef is the closest I can get, where each generation builds on the skeletons of the last and the structure rises for centuries. But a reef is not one creature. What we are watching is nearer to a single thing that hands its full height to its own next version and says start here.

The right image is not something eating itself. It is something building itself. Each generation builds the scaffold that the next one stands on, and the thing gets bigger every time round, not smaller.

Newton wrote that if he had seen further, it was by standing on the shoulders of giants. He put it in a letter to Robert Hooke in 1675, and Google Scholar still prints it under the search box. The line was not his either. It goes back to Bernard of Chartres in the twelfth century, which is the argument in miniature.

But Newton was standing on other people's shoulders. This thing stands on its own.

That is what people miss when they argue about whether the models are plateauing. The model is one ingredient. The compounding is happening in the scaffolding, and the scaffolding is being written by the thing it scaffolds.

What it looks like in my own week

I did not adopt Carson's prompts. I built his three steps into my own toolchain as two commands, so spec, tasks and execution are now part of the building rather than something I remember to do. A job starts as a few lines saying what I want and why, and the only part where I slow down is reading the spec, because that is the one place where being wrong is still cheap.

I could not have built any of it without the AI. That is not modesty, it is arithmetic.

But the reverse is just as true and people say it far less often. It could not have built this without me. Not a worse version of it. None of it.

It brings the horsepower. I bring the only thing that decides where the horsepower points, and whether what comes out the far end means anything to anybody. That is most of my job now, and it is harder work than it sounds.

The part that does not hand off

Every joint in that chain can be automated except one.

Somebody has to decide whether the spec is any good. Somebody has to notice that the whole plan is well built and pointed at the wrong thing. No amount of the machine writing instructions for the machine produces a judgement about whether this was worth building at all.

I have written before that AI is a musical instrument, and that an instrument never decides what to play. This is the same point standing further down the road. The chain can now write its own sheet music. It still cannot tell you whether anyone wants to hear the song.

Which is why I am not nervous about the fact that my editor has started finishing my half of the conversation. Of course it has. It was always going to climb that far. The grey text appears, I read it, and nine times in ten I press tab.

The whole job is the tenth time.

A free assessment, about fifteen minutes

Most businesses don’t have ten problems. They have a few parts switched off, and a few worth making stronger.
Fifteen minutes shows you which.

Contact

Tell me what you’re working on.

Say what you are trying to fix and I will give you a straight read. If it is easier to talk, pick a time that suits you.

Book a time to talk

Prefer to write?

Shei Goldberg

I read every message myself, and I will get back to you.

We will text or call you on it.