We introduce HADS (Human-AI Document Standard), a lightweight Markdown convention designed to reduce token consumption when language models query structured documents. HADS defines four semantic block types — [SPEC], [NOTE], [BUG], and [?] — and an AI manifest at the document head that enables targeted reading rather than full-document ingestion. In representative engineering documents, this reduces per-query token load from approximately 5,000 to 1,500 tokens, a 70% reduction, without any duplication of content for human readers. The standard is accessible to small models (7B parameters) without chain-of-thought reasoning about document structure. A back-of-envelope analysis suggests that at scale, HADS-formatted documentation could save on the order of 3.5 trillion tokens per day globally. The reference implementation and specification are available at https://github.com/catcam/hads.