Codebase is 250% AI generated (www.moderndescartes.com)

A developer building an AI-powered tutor for advanced chemistry reports that while Claude wrote ~98% of the lines of code, heavy human scaffolding and repeated AI-driven rewrites make the project effectively “250% AI‑generated.” The writeup argues that LLMs are astonishingly productive at churning out code, package installs, and third‑party integrations, but they need tight human direction: the single most important human task is handcrafting the information architecture (IA) — the Pydantic/SQLAlchemy models, ownership, identifiers, lifecycle and relationships — because everything else in the app flows from that DNA. Without a clear IA, the AI will “intentionlessly” design one that propagates useless features across many representations (YAML, DB tables, APIs, prompts, frontend), bloating context windows and producing brittle, inconsistent behavior. Practically, the author prescribes patterns and governance: keep core data models minimal (YAGNI), enforce consistent naming so tools like grep + LLM tool‑usage can reliably find and change code, and handwrite small framework examples so the model can emulate your style (SvelteKit components, color variables, layout rules). Tactically, use LLMs for tedious integrations, installs, and debug tooling, but pair them with human-reviewed constraints, tests, and single sources of truth to avoid drift. The piece showcases how LLMs are now first‑class engineering contributors — fast and capable, but requiring deliberate human architecture, conventions, and trimming to scale safely.