Terse plugin caps Claude Code replies at 150 words

“A 3,000-line merge gets a skim and a prayer” is the kind of sentence Claude writes when a plain one would do, and Terse is a Claude Code plugin built to stop it. According to the project’s GitHub repository, Terse cut Opus 5.5’s output on a 20-prompt test from 9,705 words to 4,468, and the run cost 32% less.

At a glance

• The author of Terse packaged 22 writing rules as a Claude Code output style, so every reply opens with the answer and stays under 150 words unless you ask for a document.
• A hook re-sends a short digest of the rules with every prompt, and on the same 20 prompts Fable 5.1 wrote 46% fewer words with the plugin than without it.
• One rule failed in practice: a ban on first person passed every benchmark, yet half the author’s real replies still said Claude would report back when the compile finished.

If you have not been following, output styles are how Claude Code changes its system prompt, the fixed instructions sent with every request. Claude Code ships four built-in styles besides Default, and Explanatory and Learning are deliberately longer than Default. According to The-Prompt-Shelf, Anthropic deprecated the whole feature on October 31, 2025 and reversed the decision four days later after users said the alternatives did not work.

The author says an engineering question gets 500 to 600 words of Claude-speak

The author calls the habit Claude-speak and lists its parts. A reply opens with “Great question” and a header, and the actual answer sits somewhere in the middle. Metaphors stand in for plain phrases, contrasts stand in for reasons (“this is intent, not behavior”), and slogans show up uninvited (“the decision is the feature”).

Then comes a recap of what Claude just did and a closing line along the lines of “happy to dig further if useful”. By the author’s count, an engineering question gets 500 to 600 words of this. Most people the author works with are tired of it, and so is the author.

Terse turns 22 writing rules into a 150-word default

Terse is 22 writing rules delivered as a forced output style. The first sentence is the answer, each sentence carries one idea, and every claim needs a number or a name. Slogans, reframes and closing offers are banned. Replies stop at 150 words unless you ask for a document.

You can install it from the Claude plugin directory by searching for “terse”, or with /plugin marketplace add lowenbjer/claude-terse. The plugin is MIT-licensed. Styles in Claude Code are otherwise picked via /config under Output style or through the outputStyle setting, since the old /output-style command was removed in v2.1.91.

Opus 5.5 dropped from 9,705 words to 4,468 on the same 20 prompts

The test ran the same 20 prompts against the same codebase, once with the plugin and once without. Opus 5.5 went from 9,705 words to 4,468 and cost 32% less. Fable 5.1 wrote 46% fewer words in the same comparison. The README also shows one prompt side by side: 636 words before, 180 after.

One rule did not survive daily use. A ban on first person passed every benchmark, yet half of the author’s real replies still said “I’ll tell you when the compile finishes”. The author’s explanation is that the benchmark never asked Claude to act as an agent. The details are in docs/benchmark.md.

A hook repeats the rules on every prompt, not only at session start

A custom output style is a markdown file with frontmatter, followed by instructions appended to the system prompt. It can live at user level in ~/.claude/output-styles, at project level in .claude/output-styles, or in a managed policy directory. The style is read once at session start, so a change takes effect only after /clear or a new session.

Terse adds a second layer. A hook re-sends a short digest of the rules with every prompt, and the author says this keeps them in force at turn 30. Think of the style as the briefing at the start of a shift and the hook as a note taped to the monitor: the first sets the rules, the second keeps them in sight.

The benchmark counts words and dollars, and the source reports no measure of whether the shorter answers stayed correct or complete. In our view, the flat 150-word cap is the rule to watch: according to Claude Code Docs, the built-in Concise style still writes at full length for error reports, failing test output, security warnings and destructive-action confirmations, while Terse’s only stated exception is a requested document.

Send the leftover Claude-speak back

If Terse still produces Claude-speak for you, the author asks you to post the reply as an issue, and the repository has a template for it. The source does not describe a fix for the failed first-person rule, and it gives no roadmap or version plan. Whether the 150-word cap holds up in agent work, where the first-person ban broke, would need a benchmark that actually asks Claude to act as an agent.