Nikolay Petrov of Intel showed that models already know roughly 70 percent of general information in agent skill files. Developers can save context tokens by removing basic concepts and retaining only exact technical parameters.