laconic is a compact skill (roughly 300 words) for AI coding agents that instructs the model to strip filler from its responses while preserving meaning, aiming for terse, context-trusting answers rather than long explanations. It is inspired by the caveman skill and targets output-token reduction specifically; it does not affect input tokens. It installs across many agent environments (Claude Code, Cursor, Copilot, Cline, Windsurf, Codex, and others) and on Pi it adds an extension with on/off toggling, natural-language triggers, a statusline, and per-project state, all reading from the same SKILL.md ruleset.
In the projects own benchmark (Claude Sonnet 4, generic system prompt vs. the laconic skill, median of 3 runs), it reports about 90% average output-token savings across ten coding tasks, ranging from 75% to 96%, compared with roughly 64% for caveman. For example, a JWT expiry question that would normally produce a full walkthrough is reduced to a one-line diagnosis plus a two-line code fix.