LLM Token Optimization for Existing Next.js AI App

We have a live AI-powered Shopify page generator (Next.js + TypeScript + Supabase, using Anthropic Claude Sonnet). Our current generation engine produces high-quality pages but consumes ~26K output tokens per page (~$0.33, ~250s). We want to drastically reduce token usage — by moving repeated CSS and page structure into code so the LLM only generates the copy — WITHOUT degrading the visual quality or copy that converts. This is optimization of an existing, working system, not a new build. You must be comfortable reading an existing complex codebase (two generation engines, streaming, structured output