In a new essay, Eugene Yan describes how to work with AI in a way that compounds over time. His core idea is that context should be treated as infrastructure: instead of pasting background into every prompt, teams should build reusable systems that supply relevant context automatically. This turns a one-off interaction into a durable asset.
He also frames taste as configuration. Rather than hoping a model knows what good looks like, users should encode their standards and preferences into the workflow. That makes quality more consistent and less dependent on luck or individual prompting style.
Yan adds that verification is the key to autonomy. Once outputs are checked automatically, people can delegate more work to AI and scale their efforts. The essay closes by emphasizing the loop: using AI, learning from results, and feeding those lessons back into context, taste, and verification. The single source does not present competing views, so the argument is presented as a unified framework.