Context capital: returns to AI co-creation and the codification of tacit knowledge. Stage 1 registration: design and hypotheses

Stage 1 preregistration; mirrored from OSF registration https://osf.io/w5uq9/ Many programmers now use AI assistants that write code for them. Some of these programmers also write a short instruction file for the assistant and store it alongside their code. The file explains how their particular project works: what to build, what to avoid, what past mistakes to steer clear of. It is a written-down version of knowledge that would otherwise stay in the programmer's head. This study asks whether the AI assistant helps more when such a file exists, and if so, which part of the file does the work. We measure how much AI assistance each public project on GitHub received, how much project-specific instruction its authors wrote, and how much durable work the project produced. We then test whether the instruction content that records human judgment (rules, warnings, corrections) matters more than content the assistant could have figured out on its own (file layouts, build commands). The project is a large-scale observational study of public software repositories.