With AI, you barely need a frontend framework

Author: dlants (dlants.me) Date: April 9, 2026 Source: https://dlants.me/vamp.html Repo: https://github.com/dlants/vamp

Summary

Argues that AI-assisted development changes the calculus for frontend architecture. Instead of React/Vue/Svelte, a lightweight vanilla-TypeScript pattern — "vamp" — is sufficient because LLMs pattern-match repetition reliably and the pattern is deliberately explicit, local, and predictable.

The View pattern

interface View<State, Message> {
  container: HTMLElement;
  sync(state: State): void;
  destroy(): void;
}

Each view manages its own DOM and state bindings. Application cycle: DOM event → dispatch → update mutates state → view.sync syncs DOM.

Key components

• Templates & bindings: Views render initial HTML using template strings with sanitize for XSS. Binder tracks state-to-DOM sync via bindText, bindAttr, bindClass, bindVisible.
• Composition: bindChild for static children, bindSlot for conditional mount/unmount, bindList for arrays with key-based reconciliation.
• Unique refs: ref() generates unique IDs to prevent selector collisions between parent/child views.
• CSS-in-JS: cls() and mountStyle() generate unique class names without build tooling.

Why this works with AI

• Prioritizes explicitness, locality, predictability over conciseness
• Repetitive patterns → AI can pattern-match and produce new views reliably
• No implicit re-renders, no hook ordering, no dependency arrays, no VDOM diffing
• Each view is self-contained → LLM reads one, understands the pattern, produces the next

React criticisms (appendix)

• Reconciliation: conditional branches remount components, losing focus/input state
• Hook complexity: dependency arrays are fragile chains; hook position must be deterministic
• Async rendering: Fiber/concurrent rendering introduces tearing, timing issues, measurement bugs
• Performance: VDOM diffing requires manual optimization (shouldUpdate, Immer)