prompt compression

Shopify explicitly achieved 4× system prompt compression; this is a direct application of prompt compression techniques

Tier-dependent compression with bounded distortion is a specific instantiation of context compression strategy

80% reduction is aggressive compression; pattern shows compression is viable without performance loss

Input optimization (reducing LLM inputs) is a form of context compression. The 2x cost savings while maintaining quality suggests effective compression.

The 'Compress' strategy directly addresses prompt/context compression through summarization.

The emphasis on tool descriptions being 'short and sweet' is a form of prompt compression—reducing context window footprint while preserving information density for LLM understanding.

Caveman-speak is a compression technique, but inverted: instead of compressing input context, it compresses output to reduce downstream token consumption. Demonstrates compression is bidirectional.

By establishing design tree structure upfront, the author is implicitly reducing context needed per phase downstream—a form of context compression through pre-work.