Thin Harness, Fat Skills: What gbrain Taught Me About Building AI Products
Most people still explain AI product progress with the wrong variable. They point to the model. Better reasoning. Better coding. Better multimodal understanding. More intelligence. But after reading Garry Tan’s Thin Harness, Fat Skills and then looking through the gbrain repo, I came away with a different conclusion. The real divide is not between products with smarter models and products with weaker ones. It is between systems that know what belongs in the harness and systems that keep stuffing more and more capability into the wrapper around the model. Garry’s argument is sharp: the harness should stay thin; the skills should be fat. The harness runs the loop, manages context, reads and writes files, and enforces safety. The skills encode judgment, process, and domain knowledge. That is where the value lives. ...