Representing LLM inputs as HTML instead of JSON to cut input tokens by 11%