For me what vastly improved the usefulness when working with big json responses was to install jq in my system and tell the llm to use jq to explore the json, instead of just trying to ingest it all together. For other things I explicitly ask it to write a script to achieve something instead of doing it directly.

That makes so much sense, it would make a great MCP. Maybe something similar for DOM manipulation; extracting text out of big, noisy HTML pages using a combination of Find Text with selector return values, and a DSL for picking and filtering DOM trees.