Pure Code: Preserving Judgment, Understanding, and Experience in AI-Assisted Development

Summary

Pure Code is an essay’s proposed approach to AI-assisted development: a contribution should preserve the purpose of a change, its design decisions, verification, limits, and the contributor’s ability to engage with future work. The proposal is framed by concerns about AI-generated open-source submissions that leave maintainers with unusable patches, excessive changes, and little useful follow-up, while remaining distinct from Zig’s rules, which prohibit LLM-generated code and prose. The essay argues that faster implementation is not automatically sustainable if it shifts unexplained investigation and correction to maintainers. It contrasts two contributors who use AI to implement the same persistence fix: one forwards a patch without understanding it, while the other records the relevant distinction between interface updates and saved state, checks recovery after restart, and identifies untested conditions. Pure Code would begin with a contribution record covering purpose, design choices, AI involvement and human review, exact verification steps and results, limits, and continuity through review and handoff. The record is intended as evidence of what happened, not proof that a person understands the change. The essay says AI-generated explanations and passing-test claims can be misleading, so human decisions, model interpretations, executed checks, and observed review responses should remain distinguishable. It positions the idea alongside LLVM’s AI policy, SLSA provenance, and the Developer Certificate of Origin, while emphasizing continuity of experience rather than replacing those practices. Cleanroom is described as one implementation, but the essay presents no measured evidence that the approach improves learning or reduces maintenance costs. It proposes evaluating maintainer effort, repeated review issues, post-merge defects, contributor reasoning, and whether project knowledge is reused in later work.