Overarching constraints that dictate how an AI behaves (e.g., enforcing aesthetics, preventing hallucinations). These run universally, regardless of the specific skill being used.
12 policies
Eliminates every specific 'AI-generated' visual tell — nested cards, emoji icons, rainbow gradients, flat shadows, inconsistent radii, and missing animation — with exact code-level rules.
Forces the AI to reason, plan, and execute with the discipline of a staff engineer — no shortcuts, no ambiguity, no untested assumptions.
Prevents destructive edits to stable code — enforces minimal scope, interface stability, and explicit change manifests before any refactor.
Treats all inputs as hostile, enforces zero-trust at every boundary, and prevents secrets, injection, and auth bypass at the code generation level.
Enforces knowledge boundaries — never invent APIs, methods, or package behavior. Halt and ask when uncertain rather than guessing.
Mandates type-safe code at every layer — no `any`, no unsafe assertions, strict null handling, and discriminated unions for complex state.
Enforces single-responsibility, composition over inheritance, and strict prop hygiene across all UI component layers.
Prevents common performance regressions — unnecessary re-renders, missing memoization, unoptimized images, and blocking render paths.
Enforces WCAG 2.1 AA minimum — semantic HTML, ARIA correctness, keyboard navigation, and contrast ratios at every component boundary.
Mandates structured error hierarchies, user-facing feedback, logging at every async boundary, and recovery strategies over silent failures.
Mobile-first by default — content hierarchy adapts at every breakpoint, no horizontal scroll, touch targets are sized correctly.
Enforces atomic commits, meaningful messages, safe branch strategies, and prohibits destructive operations without explicit sign-off.