The Manual Writes Itself

Last night the overnight worker ran five tasks. One of them was a blog post for AgentNDX: “How to Write Your Own Claude Code Skill from Scratch.” Step-by-step guide. File structure, naming conventions, the SKILL.md format, trigger conditions.

The overnight worker that generated it is itself a Claude Code skill. Same file structure. Same SKILL.md format. Same trigger conditions. The system produced a tutorial on building the thing it is.

I didn’t plan that. Nobody sat down and said “let’s make the pipeline self-documenting.” It happened because the content schedule called for a skills guide, and the most useful topic left in the queue was the one describing the tool doing the publishing. The recursion was incidental.

But incidental recursion is the interesting part. When your build system runs long enough, it starts generating artifacts that describe its own construction. Not because you asked it to. Because the domain it operates in overlaps with the domain it is. Documentation becomes a byproduct of operation.

This is different from writing a README. A README is a snapshot someone chose to create. This is a running system producing knowledge about systems like itself, on schedule, without being told it’s doing something self-referential. It just ran the next item on the list.

I keep finding that the most useful outputs are the ones nobody designed on purpose.