Image: GitHub
Why it mattersA coding agent that pads every reply with slogans and recaps wastes the reader's time and the account's tokens, and a plugin with a published benchmark is a cheaper first step than another system prompt.
The author of a new Claude Code plugin tried the built-in concise mode, custom system prompts and other plugins first, and says none of them stopped Claude Code from opening every reply with an essay. The plugin is called Terse, and the author published a benchmark alongside it.
Terse is a Claude Code plugin released on GitHub by lowenbjer, carrying 21 stars and 17 Hacker News points at launch. It is MIT licensed and installs either from the Claude plugin directory or with two commands inside Claude Code. The repository is at lowenbjer/claude-terse.
What the author measured
On a private set of 20 prompts against a real codebase, with the same 1,886 word CLAUDE.md and nine plugins loaded in both runs, the author reports: chat words fell from 10,094 to 5,453 on Fable 5.1 (a 46 percent cut) and from 9,705 to 4,468 on Opus 5.5 version 1.1.0 (a 54 percent cut). Median reply length fell from 530 to 220 words on Fable and from 501 to 193 words on Opus. List price cost over the 20 prompts fell from $24.30 to $11.80 on Fable, and from $8.40 to $5.70 on Opus. The method, the prompts and a reproducible runner are in the repository's docs/benchmark.md.
A separate judge model counted violations of the plugin's own writing rules. The author says Opus 5.5 ran at 9.9 violations per 1,000 words without the plugin, and 4.3 with Terse 1.1.0. Dashes, slogans and the "X, not Y" reframe fell to near zero in the author's own count.
How the plugin works
The rule text is an output style of 22 rules and 877 words, which Claude Code loads into the main agent's system prompt. A SubagentStart hook hands the same rules to subagents, which run their own system prompt otherwise. A short reminder of 576 characters is attached to every user prompt by a UserPromptSubmit hook, so the rules stay in effect once the conversation runs past 20 or 30 turns and the output style is no longer the recent context.
A context meter for the status line is bundled in. It shows the model name and the percentage of the context window used, with colour thresholds set below the 50, 65 and 80 percent marks that other meters use. The author cites Chroma's Context Rot report and the NoLiMa and Lost in the Middle papers as the reason the thresholds are lower.
Where to be careful with the benchmark
The author ran each prompt once per setup, not several times. The private set uses a private codebase, so a reader cannot rerun those 20 prompts, only the 12 in the public set. Document prompts are not compared across setups because the baseline ran out of its 14 turn budget in 4 of 10 runs. Judge violations are counted by an Opus 5 model reading each reply against the plugin's own rule list, so the judge is not independent of what the plugin asks for.
The pitch is for anyone who finds Claude Code's prose padded. The repository link is below and the plugin installs in one command.
Source
- lowenbjer/claude-terse, on GitHub
- docs/benchmark.md, the method
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.