Show HN: Graft – Claude Code hooks that cut grep tokens by 42%

There's something almost paradoxical about optimizing an AI's reading habits. We spend so much time building models that can process entire libraries of text, only to realize that their true bottleneck isn't intelligence—it's attention. Graft doesn't make Claude smarter; it makes Claude *lazier* in the most productive way possible. By filtering grep results before they ever hit the context window, it forces the agent to work with surgical precision instead of brute-force recall. This is the digital equivalent of teaching a supercomputer to use a highlighter instead of photocopying the whole book. The 42% figure is deliciously evocative, whispering of Douglas Adams and the search for the ultimate answer. But here, the answer isn't 42—it's efficiency. In the emerging economy of tokens, where every API call costs real money and every context window has a hard ceiling, Graft represents a new class of "meta-tools" that don't just write code, but manage the *cost of thinking*. It's a glimpse into a future where we don't just optimize algorithms, but optimize the very resource allocation of machine cognition. This also speaks to a deeper, stranger truth about human-AI collaboration. We're moving past the era of "give the AI everything and let it figure it out" into an era of "teach the AI to be selective." It's a form of cognitive pruning, a digital Zen garden where less is genuinely more. The wildest part? We're building these constraints based on our own human intuitions about what matters—passing our biological biases about relevance into the silicon mind. Graft isn't just a tool; it's a philosophical statement that the future of intelligence, both natural and artificial, lies not in raw capacity, but in the elegance of what we choose to ignore. Source: [https://github.com/NanoNets/Graft](https://github.com/NanoNets/Graft)