Ask HN: What is a good format for a tool to report data to a LLM?
TL;DR
With CLI tools the picture is clear: humans want cleanly formatted text, sometimes with colors or a TUI, while machines prefer tab-separated tables or JSON. For LLMs there is no comparable rule of thumb yet. A Hacker News thread asks exactly that question: which format helps a model grasp tabular or structured data quickly and reliably? The discussion is still small, but it hits a point familiar to anyone building tools for agents.
Nauti's Take
The question points at a real advantage: choosing a tool's output format deliberately cuts token cost and error rates in a measurable way. The catch is that solid benchmarking barely exists here, mostly gut feel from individual projects.
Anyone building agents should test two formats against each other rather than trust a single forum recommendation.
Sources