Kills self-repetition in LLM responses. Cuts 30-40% wasted tokens by eliminating restated questions, echoed instructions, and repeated ideas.