What problem does it solve? LLM agents waste large amounts of context tokens narrating filesystem facts, git state, and verbose build/test output that the terminal can answer exactly and cheaply. This Skill substitutes deterministic shell commands for LLM reasoning about facts and clamps command output by 60-90% before it reaches context. ## Core Features & Use Cases - Terminal-first substitution table: Cross-platform commands (git, gh, rg, python3) for file existence, branch state, PR lookups, disk space, and JSON extraction, so facts come from the shell instead of LLM guesses. - Output-reduction catalog: Per-command clamp recipes for test runners, type checks, diffs, lints, and commits, with signal-tiered truncation caps that always preserve error lines. - Failure-safe tee cache: On non-zero exits, full unfiltered output is written to a log file and only the path is surfaced, with a retrieve convention to fetch the original on demand. - Use Case: While running a CI fix loop, an agent clamps a raw tsc --noEmit output from roughly 2000 tokens down to about 80 tokens of error lines, keeping the full log recoverable via the tee cache if more context is needed. ## Quick Start Ask the agent to check the current git branch and run the test suite using terminal-first commands with clamped output instead of narrating the results.
