tldr

Your coding agent buries the answer in four paragraphs. tldr makes it lead with a three-line TL;DR and fold the full detail underneath. The rule is demote, don't delete — compression moves information down the page, it never removes it. It also fixes the expensive version of the problem: a subagent handing the orchestrator 4,000 tokens of narration where three lines would do. You pay for that twice — once to write it, once to carry it in context. Most "be concise" prompts save tokens by deleting information, which is a bad trade during a code review or a security audit. tldr has an explicit never-compress list: destructive actions and their blast radius, security findings, anything that bills, and verbatim error text always stay above the fold. One paste installs it in 20 agents — Claude Code, Cursor, Codex, Gemini CLI, OpenCode, Zed, Windsurf, Cline, Aider and more. No runtime, nothing to compile, MIT. Measured, not asserted: 16 cases x 3 trials, blind-graded against a no-skill baseline on two models. Agent-to-agent output falls 37% on Sonnet and 59% on Opus, and it scores higher on correctness, fidelity, actionability, safety and concision on both. All thirteen eval runs are published, including the four that failed.