Terse plugin caps Claude Code replies at 150 words
Terse, a Claude Code plugin its author showed on Hacker News on October 6, cut Opus 5.5’s output on 20 prompts against one codebase, run with and without it, from 9,705 words to 4,468 and its cost by 32%, by the author’s own benchmark in the project’s GitHub repository. With the plugin enabled, its 22 writing rules apply as a forced output style that limits a reply to 150 words unless a document is requested.
Installation runs through the Claude plugin directory (search “terse”) or /plugin marketplace add lowenbjer/claude-terse, and the license is MIT. Claude Code’s documentation says a style that a plugin forces overrides the user’s own outputStyle setting whenever the plugin is enabled, so turning Terse on replaces whatever style was picked, including Concise, the built-in style Anthropic added in v2.1.237 on August 19 that states results first and drops preambles.
Claude-speak, as the author calls it, opens with “Great question” and a header, hides the answer in the middle, reaches for a metaphor where a plain phrase would do (“a 3,000-line merge gets a skim and a prayer”), drops a slogan, recaps what it just did and ends with “happy to dig further if useful.” An engineering question gets 500 to 600 words of it, the author says, and the README’s example reply goes from 636 words to 180.
Terse’s rules run the other way: the first sentence is the answer, one idea per sentence, a number or a name in every claim, and no slogans, reframes or closing offers. A hook re-sends a short digest of them with every prompt, so the rules still hold at turn 30.
Fable 5.1 wrote 46% fewer words on the same prompts, and on Terse 1.0.0 its list-price cost fell 51%, from $24.3 to $11.8. Both baselines already carried the author’s own 1,886-word writing rules in CLAUDE.md, so Terse’s savings come on top of them. Each prompt ran only once, through claude -p, in September. The author’s private prompt set, adapted from real sessions against a private codebase, is published in aggregate, while a public set of 12 prompts can be rerun with bench/run.py.
A ban on first person passed every benchmark and lost in daily use: half the author’s real replies still said “I’ll tell you when the compile finishes.” The benchmark never asked Claude to act as an agent.
Before choosing the plugin text, the author ran the same set against eight rule texts and placements, with Opus 5 grading each reply on a nine-item rubric: wording and placement moved the judged categories by 15 to 50 percent, and metaphor did not move in any configuration.
“If it still writes Claude-speak, post the reply as an issue,” the author writes; the repo has a template for it.
Up next
Anthropic’s daily-brief agent treats a failed read as news
One stale “still waiting on you” costs more trust than ten missing items, Anthropic writes in its Claude Managed Agents guide, so its scheduled daily brief re-checks every item right before posting to Slack.
Related stories
- Chisle’s author says it cuts oversized tool output by 48%
- Every Claude Code system prompt is now public, tracked across 238 versions
- Tokendoctor cut its author’s Claude Code cost 38%
- Claude Code Ultra week measured at 17B Opus cache reads
- memsaga gives Claude Code’s memory files a git history
- Intent-lens spells out Claude Code calls before they run
Comments
No comments yet. Be the first.
Join the conversation
Sign in to leave a comment. It appears after moderation.

