Lab

An open-source Claude/Codex skill that forces a research pass — checking prior art, constraints, and tradeoffs — before proposing how to build something. Built to stop AI-assisted architecture advice from being confident and wrong.

Encodes four principles distilled from Andrej Karpathy's observations on LLM coding failures: think before coding, simplicity first, surgical changes only, and goal-driven execution with verifiable success criteria. One file, drop into any project.

A portable skill that upgrades layout, typography, motion, and spacing in AI-generated interfaces, with adjustable dials for design variance, motion intensity, and visual density. Stops output from looking like default AI slop.

Covers OWASP Top 10:2025, ASVS 5.0, the LLM/RAG-specific Top 10, and 2026 Agentic AI security risks, plus language-specific vulnerability patterns for 20+ languages. Activates automatically on auth, input handling, and API work.