How I strip 90% of code before feeding it to my coding agent

Context windows keep growing. 200k tokens. A million. The assumption is that bigger context means better answers when working with code. It doesn't. The attention problem Take a typical 80-file TypeScript project: 63,000 tokens. Modern models handle that easily. But context capacity isn't the bottleneck — attention is. Research consistently shows that attention quality degrades in long contexts. Past a threshold, adding more tokens makes outputs worse. The model loses track of critical details, latency increases, and reasoning quality drops. This is the inverse scaling problem: more context, worse outputs. When you ask an AI to explain your authentication flow or review your service architecture, it doesn't need to see every loop body, error handler, and validation chain. That's 80% of your tokens contributing nothing to the answer. What signal actually matters For architectural understanding, the model needs: What functions and methods exist (names, parameters, return types) What type