The getting started guide was one sentence.
Jay found it in the Attractor README, in the section labeled "Quick Start." Below the installation instructions, below the dependency list, below the architecture overview, there was a single line of text:
Supply the following prompt to a modern coding agent: Implement Attractor as described by factory.strongdm.ai.
That was it. That was the whole getting started guide.
Jay read it three times. Then he laughed—not the polite laugh of someone acknowledging a joke, but the involuntary laugh of someone who has just seen something perfectly self-referential.
"Justin," he said. "Did you write the Quick Start section of the Attractor README?"
"Yes."
"The getting started guide for the AI software factory is: tell an AI to build the AI software factory."
"Yes."
"That's... recursive."
"That's the point." Justin set down his coffee. "If the methodology works, it should be able to produce itself. The factory builds software from specifications. Attractor is software. Its specifications are public. Therefore, a coding agent should be able to build Attractor from its specifications. If it can't, the methodology has a gap."
Navan, who had been listening, opened a terminal. "I want to try it."
"Right now?"
"Right now."
He opened Claude Code in a fresh session, inside a Leash container because of course inside a Leash container, and typed the prompt exactly as the README specified. Then he sat back and watched.
The agent read the specifications on factory.strongdm.ai. It parsed the attractor-spec.md, the coding-agent-loop-spec.md, and the unified-llm-spec.md. It built a mental model of the pipeline runner—the DOT graph parser, the node handlers, the edge-based routing, the checkpoint system. Then it started generating code.
Jay and Justin watched over Navan's shoulders. The agent worked for forty-seven minutes. It produced a Go implementation of the Attractor pipeline runner. It implemented DOT graph parsing, node execution with pluggable handlers, edge routing with conditions and weights, checkpoint and resume, and the event stream that drove the TUI frontend.
Navan ran the scenarios.
Satisfaction: 91%.
"On the first attempt," Jay said. "From a one-sentence getting started guide."
"The guide isn't one sentence," Justin corrected. "The guide is 5,700 lines of specification. The README just tells you where to point the agent. The specs do the rest."
"The spec is the documentation is the product is the getting started guide," Navan said, and it sounded like a koan.
"It's not a koan," Justin said, as if reading Navan's tone. "It's just... consistent. The factory is built on the premise that specifications are sufficient for implementation. The README demonstrates that premise. One sentence pointing to the specs. If you need more than that, the methodology needs work."
Navan bookmarked the satisfaction score. Ninety-one percent on a cold start from a public specification. The factory could build itself.
The README stayed one sentence. It didn't need to be longer.
A software factory whose getting started guide is "tell an AI to build this software factory." The self-referential loop is perfect. If the methodology can't produce itself, it doesn't work. If it can, the README is sufficient.