typingtube's 200 plans grow with zero 'write a plan' orders

The article frames the proliferation of plans as the mirror image of a rule stated in an earlier installment of the series: rules handed to an AI do not grow as you move up the hierarchy, but plans do. On typingtube, they have grown to 200, one per feature, despite the developer's explicit finding that 0 numbered procedures exist among them. The mechanism the author describes is precedent: the AI reads existing plans before starting work, matches their format, and takes the next number, so the second plan was summoned by the first rather than by a request.

The plans are not purely mechanical output. Across 73 of them, the phrases 'user confirmed,' 'user decision,' and 'user instruction' appear 538 times, and the developer says he still writes those judgments himself. What he stopped was asking for a plan, not deciding one. The distinction matters because it separates two things that are usually bundled: the request that a plan be produced, and the content of the plan.

The author flags three limits he has not measured. The very first plan needs a person, and behavior with zero precedent was never tested. One of the two sub-agents also cited a leftover placement line in the instruction file, and the effect of deleting that line is unknown. Whether this pattern holds outside a repository with a few plans already in place is, on the article's own account, an open question.