as fun as it is to get claude to speak in a more token efficient language (caveman, chinese, etc) -- using actually good connectors/mcps and setting up basic progressive disclosure within your documents/skills is going to get you a lot farther and save a lot more $$$

expectation: caveman mode has reduced response output tokens by xx%!

reality: the glean mcp dumped a 20KB payload directly in the tool response and the agent spent $10 just figuring out how to read it w/o killing its context window