If you've ever asked Claude a quick question and gotten back a five-paragraph essay, you're not imagining it. The Claude Terse plugin is a small add-on for Claude Code, built by one developer who got tired of reading essays and decided to fix the problem himself.
What happened
A developer who goes by the handle lowenbjer published a plugin called "terse" on GitHub, under the repo name lowenbjer/claude-terse. His pitch is blunt: no more essays, no more slogans and metaphors, no more "it's not X, it's Y." The plugin makes Claude speak normally.
That's the whole idea. He tried system prompts first — the standing instructions you give an AI to set its tone — but they got forgotten after a couple of turns. He tried other plugins, but each one solved a different problem: cutting token costs, or making output easier on neurodivergent readers. None of them simply made Claude get to the point. So he spent a few weeks writing down every habit that annoyed him, plus the phrasing he wanted instead, and turned that list into rules.
The result is a Claude Code plugin that roughly halves reply length and strips out the flowery prose. The repo has 15 commits so far. Even the README is partly written by Claude Code running the plugin — the author jokes that he takes most of the credit but none of the words.
The benchmark numbers, in plain English
Two sets of tests are published. Across 20 prompts run against a real codebase, Fable 5.1 produced 46% fewer words and cost 51% less, while Opus 5.5 produced 54% fewer words and cost 32% less.
A second test compared plain Claude Code — no plugin, no custom instructions — against terse 1.1.0 on 12 public prompts, using Opus 5.5 at high effort, one run each. The results are laid out in two columns: vanilla, meaning Claude Code with no CLAUDE.md file, no plugins and no custom instructions, and terse. Here's what moved:
• Across 10 chat replies, the word count fell 70%.
• The median reply fell from 636 words to 181 — down 72%.
• Output tokens dropped 45%.
• Style violations per 1,000 words fell 90%.
• Slogans and metaphors per 1,000 words both dropped to zero.
What counts as "slop," exactly
The rules go after specific habits. A slogan is a short line that sounds like a principle but explains nothing — "The decision is the feature." A metaphor is figurative wording where a plain phrase would do, like calling a 3,000-line merge "a skim and a prayer." An "X, not Y" reframe uses a contrast as a label instead of an explanation — "This is intent, not behavior." Cadence is rhythm built for effect, such as triads and punchline endings. Bloat is a sentence that adds no fact, like "I'll look at the code now." And first person means sentences whose subject is the writer: "I," "me," "my," "let me."
What it means for you
You don't have to be a programmer for this to matter. If you use Claude Code for anything at all, here's what changes.
At home
You ask Claude to sort out a messy folder of files or fix a script on your laptop. Instead of three paragraphs of throat-clearing, you get the diagnosis and the steps. Shorter answers are also easier to follow on a phone screen when you're not sitting at a desk.
At work
If part of your job is reviewing what an AI wrote, terse output cuts your reading time hard. The measured median reply dropped from 636 words to 181. Multiply that across a daily back-and-forth and you're getting an hour back per week, not 30 seconds.
Running a business
Fewer output tokens is a smaller bill. Output tokens — everything the model writes, thinking included — are priced about five times higher than input tokens on Opus and Fable. A 45% cut in output tokens shows up on your invoice. If you want other free ways to shave time off your workflow, there are handy free AI tools worth bookmarking alongside this plugin.
Studying
A terse reply is far easier to turn into flashcards or notes, because it's already stripped of filler. You're copying facts and steps, not untangling rhetorical flourishes.
Creative work
Here's the honest catch: if you want vivid prose, this plugin is not your friend. It kills metaphor on purpose — measured at zero per 1,000 words. Use it for editing, planning and outlines, and turn it off when you actually want voice.
Earning extra income
People who sell AI-assisted services live and die on turnaround time. A shorter, plainer reply takes less time to quality-check before it goes to a client, which means you can take on more work with the same hours.
How to try it right now
The plugin is free — it lives in a public GitHub repo, so there's nothing to buy beyond whatever Claude access you already have. Here's the path, in order:
1. Confirm you have Claude Code running. This is the terminal-based version of Claude that works inside your project folders, not the regular chat app.
2. Find the repo. Search GitHub for the project named lowenbjer/claude-terse — the plugin itself is called terse.
3. Follow the install steps in the README. The usual pattern in Claude Code is `/plugin marketplace add lowenbjer/claude-terse`, then an install command for the plugin. Copy the exact commands from the README, since plugin syntax changes.
4. Restart Claude Code so the plugin loads.
5. Ask it something real — a bug fix, a summary, a plan. Then look at how many lines came back.
6. Check your status line. A context meter ships with the plugin and shows how full your context window is, so you can see when the conversation is getting crowded.
The writing rules apply to more than chat replies. They also cover documents, commit messages (the short notes developers attach to code changes) and subagents — helper instances Claude spins up to handle side tasks on its own.
Upsides and what changes
The big win is durability. A system prompt is a suggestion the model can drift away from after a few exchanges; a plugin applies the same rules every turn. The published numbers back that up: style violations per 1,000 words fell 90%, and first-person phrasing dropped 83%.
The cost angle is real too. Cutting output tokens by 45% on work where output is priced roughly five times higher than input adds up over a month. You also get the context meter, which most people end up using more than they expect.
Limitations
The benchmark is small and the author says so. It's one run per prompt, the judge is an Opus 5 model counting violations against a rule list, and most of the testing happened on code work — so your mileage will vary by task. Terse also doesn't make Claude smarter or more accurate; it only changes how it talks. If you enjoy the personality, the jokes and the flourishes, you'll find this plugin flat and robotic. And it's built for Claude Code specifically, not every place you might talk to Claude.
Conclusion
One developer got annoyed enough at AI padding to build a plugin, and the measured result is roughly 70% fewer words in chat replies with zero slogans and zero metaphors. You don't have to take the concept on faith — the numbers are in the open.
Your one action today: open the repo, install terse in Claude Code, and run the same prompt you ran yesterday. Count the words in both answers. If the second one is a third of the length and still says everything you needed, leave it on.
Comments 0