Tokeniser-Aware Shorthand (TASS): A Stenography-Inspired Output Format for Reducing LLM Inference Cost in Structured Extraction Pipelines

Output serialisation formats designed for human readability—JSON, YAML, even the newer TOON specification—carry structural overhead that is purely an artefact of human-facing conventions. When the consumer is a large language model and the producer is another large language model, those conventions tax every inference call. This paper proposes a stenography-inspired shorthand notation aligned to LLM tokeniser boundaries, achieving a 75–85% reduction in output tokens versus JSON for fixed-schema structured extraction workloads, with parsing complexity equivalent to a single split-and-map operation on the server. The approach is validated against a production rate calculator use case but generalises to any production system performing high-volume, fixed-schema LLM data extraction.