newsfilter
← Interviews
AI Engineer14 min

Your SKILL.md Is 70% Stuff the Model Already Knows - Nikolay Petrov, Intel

Nikolay Petrov

Conference Presentation

The event details a rigorous methodology for defining and evaluating AI skills by decomposing them into discrete atoms tested against four capability quadrants. Through strict evaluation of models like Anthropic Opus and Sonnet, researchers identified that while larger models generalize well, smaller variants face degradation risks when skills introduce instruction interference. Optimization strategies leveraging the docume framework then refine these skills by pruning redundant information, reducing context costs to approximately $14 per evaluation cycle.