Claude Code First, specification version 0.1.0. The plain-language guide

Version 0.1.0. Grade R3/S3.

F1 An every-turn instruction file grows until a gate refuses the next addition, and a budget stated in prose did not stop one

FieldValue
Statusactive
TopicContext economy and agent prose
GradeR3/S3
CasesShio, roadkeep
Contrarynone
Harness factsH1, H2

Claim

An instruction file loaded on every turn grows as work is added to the project: in Shio it grew about nineteenfold under a token-budget law stated in prose, and in roadkeep it grew to the ceiling of a gated budget and stopped there, after which content had to be compressed or moved out.

Mechanism

The every-turn file is where a finished task's rationale is cheapest to write down, so each task tends to add to it and none removes from it. A budget stated in the same file is read by the agent that is adding to it and checked by nothing, so it does not refuse the addition. A budget held by a gate refuses the addition that would cross it, which turns the next addition into a decision about what to move or cut. The file loads on every turn (H2), including what it imports (H1), so its size is paid on every turn of every session.

Observations

Independence

Two independent origins. Shio's growth happened in Shio's file. roadkeep's growth to its ceiling happened in roadkeep's own agents.md, a different file with a different history, so it is a second occurrence of growth even though roadkeep's gate was designed after Shio's failure. The gate itself is one remedy with one origin, roadkeep (grading rule 2); Shio does not carry it for agents.md. The comparison between a prose budget and a gated one rests on one case of each: the prose budget failing is seen only in Shio.

Counter-evidence

Grade

R3: growth of the every-turn file is observed independently in two projects. S3: the strongest observations are counted, Shio's byte sizes at four commits and roadkeep's own readings of 125 of 125 lines and then 104 of 125. Not S4: the budget check exists and encodes the claim, but no source shows it failing. A rule about prose budgets alone rests on the Shio observation, which is R1/S3.