Shopify's engineering introduced gisting, a novel technique for compressing long LLM prompts into a smaller set of learned "gist" tokens, improving throughput and reducing inference cost.