Reduce unnecessary context, file reads, searches, tool output, and response length while preserving correctness, safety, required verification, and user intent. Use when the user explicitly requests token-efficient execution in ChatGPT or Codex; never use it to justify skipping necessary evidence.
