Laconic

Laconic is a 300-word skill for AI coding agents that instructs the model to answer tersely: cut filler, keep meaning, trust the reader's context, and emit only as much code as the answer needs. It is inspired by the caveman skill but aims for a terse rather than primitive register. The skill only reduces output tokens; input context is unaffected.

In a benchmark of 10 coding prompts run on Claude Sonnet 4 through the Langdock API (median of 3 runs, generic system prompt versus the full laconic SKILL.md), average output fell from 1,547 to 139 tokens, a 90% reduction with a range of 75% to 96% per task, compared with an average of 64% for caveman on its own benchmark. The benchmark script is included and can be rerun with an API key.

The skill installs through the skills CLI for Claude Code, GitHub Copilot, Cursor, Windsurf, Cline, and 40+ other agents, as a Claude Code plugin via a marketplace, as a Codex plugin, or as a pi package. On pi it also ships an extension with /laconic [on|off], natural-language triggers, a status line, and per-project state in .pi/laconic-mode.json.

Features

• Brevity skill: a ~300-word SKILL.md that trims explanations, hedging, and boilerplate from agent replies
• Output-only savings: reduces generated tokens; does not change how much context is read
• Published benchmark: 10 tasks, 90% average output-token reduction, reproducible with benchmarks/run.py
• Broad install paths: npx skills add, Claude Code plugin marketplace, Codex plugins, and pi packages
• pi extension: /laconic toggle, phrase triggers such as "be laconic" and "normal mode", and a status-line indicator
• Trigger phrases: /laconic, "laconic mode", or "be laconic" activate the skill in a session