Asking an AI to use verifiable info changes its picks dramatically

santiviquez · x · 2026-11-17

While testing tools, the author found that simply asking the model to rely on verifiable information significantly changed its recommendations. He reflects that output quality still depends heavily on how much the person prompting knows the right questions to ask. Having spent too much to quit, he pivoted to a new approach: having an agent read each tool's API documentation and verify whether it could handle each of 10 specific use cases, to find which tool is actually best.

Original post →

More from coding & agent

coding & agent channel →