Cache-Prefix-Change-Suffix Principle

Description

Convergent principle: split every prompt into a static cacheable prefix and a dynamic session-specific suffix. Claude Code uses SYSTEM_PROMPT_DYNAMIC_BOUNDARY marker to cut inference cost ~80% on long Opus sessions. AutoBE's executeCachedBatch establishes cache with first task then parallelizes the rest at 90%+ cache hit rate, ~88% token cost reduction on 40-API batches.

Key claims

Relations

Sources

src-20260419-16b155f4f619