less

A communication protocol between you and Claude.

Claude info-dumps by default. less makes it speak like someone who respects your attention: answer first, keep every fact that matters, cut the rest, offer choices as a pick-list you can answer with one letter.

It shapes only what Claude says to you — never how hard it thinks, what tools it runs, or how other plugins behave. Less noise, not less signal. Not a token-saver; a mental-space saver.

What it does

• Answer first. The result or decision is line one. No preamble, no recap, no "Great question", no sign-off pleasantries.
• Keeps the load-bearing facts. Numbers, paths, risks, breaking changes, anything irreversible. Minimal ≠ missing.
• Breathes. One idea per line. Whitespace over walls.
• Selectable choices. Decisions render as the native interactive picker (arrow-keys + enter), recommended first — you select, you don't type. Clean UI is part of the calm, not a separate thing.

Modes

command	behavior
/less:on	default — minimal, answer-first, pick-lists
/less:zen	near-silent — results and pick-lists only
/less:off	normal Claude verbosity
/less:less	show the current mode + the switches

Use the namespaced form (/less:zen, not /less zen). A bare /less zen collides with Claude Code's skill-stacking parser (2.1.199+) and errors with args from unknown skill: zen. Modes take effect from the next turn.

Install

claude plugin marketplace add nagisanzenin/less
claude plugin install less@less

Active from the next session.

How it works

A UserPromptSubmit hook injects the protocol (protocol.md) into Claude's context each turn — silent to you, additive to everything else. It replaces nothing, so it can't degrade Claude's work or collide with other plugins. Turn it off and Claude is exactly its normal self.

Uninstall

claude plugin uninstall less@less

More from the same workshop

Five Claude Code plugins from the same workshop. Most share one habit: let a deterministic core decide, and never let the producer of work grade it.

• effortmining — benchmark-calibrated per-subagent reasoning effort: dispatch the cheapest tier a blind grader still accepts. ~64.7% fewer output tokens at equal quality, pre-registered.
• engram — evidence-based learning engine: first-principles curricula, generation-first tutoring, blind-graded free recall, FSRS-scheduled memory. Learn anything; keep it.
• idiolect — human-voice writing engine: 60+ measured voices plus a deterministic AI-tell scanner and a blind auditor, so text reads like a person, not a model.
• production-grade — turns "build me X" into a gated multi-agent pipeline (architecture → tests → security → CI/CD) with a receipt for every phase.

License

MIT