Show HN: Terse, a Claude Code plugin that halves reply length by cutting filler

Terse is a Claude Code plugin that enforces a strict concise writing style for assistant replies, documents, commits and subagents. It ships as an output style and a 22-rule rulebook that prescribes a reply shape: one sentence with the answer, then a list or table, then at most one caveat sentence, then stop. The rules ban first-person framing, metaphors, "X, not Y" reframes, slogans, em dashes, self-labeling, bloated sentences and other mannered prose; bullets must start with facts and sentences must use the topic as the subject. The rule set is injected into every prompt as a short reminder (≈145 input tokens per turn), subagents are handed the same rules on start, and a separate context-meter script shows model name and context fill with thresholds that turn green/yellow/orange/skull at 37%, 49% and 60%.

Benchmarks against vanilla Claude Code show large reductions in verbosity and style violations: measured on Fable 5.1 and Opus 5.5 across real prompts, terse produced 46-54% fewer words and lowered output tokens by ~32-45%, with median reply shrinking (e.g., Opus median 636 → 181 words). An Opus judge found style violations per 1k words drop ≈90% and first-person uses drop ≈83%; metaphors and slogans went to zero in tests. Install via the Claude plugin marketplace or CLI; a one-time /terse:install-meter adds the meter to the status line.