Jet Xu writes about DocMason, local-first knowledge bases, and the systems needed to turn messy private files into reliable, evidence-backed context for serious work.
Two blind copywriting tests for a project called CodexWork showed that AI slop is rarely about bad English. It is a failure to read where the visitor is standing: which page they are on, what they already know, and what the next click should feel like.
Generic AI reviewers fail where enterprise context begins. The next valuable AI code reviewer will behave like an internal digital staff engineer: platform-aware, company-aware, human-looped, and accountable for merge trust.
The AI industry's pursuit of a single omnipotent God Model is a dead end. Due to the mathematical paradoxes of RLHF and alignment, models are inevitably bifurcating into specialized hemispheres—and the future belongs to Agentic Frontal Lobe orchestration.
In early 2025 I started repo-graph-rag to build token-free, traversable repository intelligence. The recent code-graph wave validates the problem, but also explains why I moved upstream anyway.
After building DocMason with AI agents, I stopped treating vibe coding as prompt craft. The durable layer is repository governance: Dao, Fa, Qi, Shu, AGENTS.md, skills, planning, verification, and closeout.