
sharur
Cut Claude Code token usage and cost with a one-line output style. Shorter replies than Concise on every Claude model: 29% fewer output tokens and 23% lower…
Install with your AI
Paste into Claude Code, Cursor, or any agent — it reads the repo and wires the tool into your project.
Install and set up sharur (git-clone project) into my current project. Found on https://claudeers.com/sharur Repo: https://github.com/ilien-dev/sharur Homepage/docs: — Detected install method: git-clone → git clone https://github.com/ilien-dev/sharur Category: devtools. Platforms: cli. Read the repo's README for exact setup and env vars, then install it and wire it into my project. Claudeers Health Verdict: unknown; community-verified: false. Confirm the source before running anything.
git clone https://github.com/ilien-dev/sharur
// compatibility
| Platforms | cli |
|---|---|
| Operating systems | — |
| AI compatibility | claude |
| License | MIT |
| Pricing | open-source |
| Language | Python |
English · Español
Sharur: cut Claude Code token usage with a one-paragraph output style
An output style for Claude Code that answers in one or two lines, with no greeting or filler,
and one short line per item when the result is bigger. One short paragraph in one small file.
Install · What you save · Before and after · vs caveman, rtk, ponytail · Before you switch · Benchmarks
[!TIP] On everyday coding tasks, Sharur costs 21% less than Claude Code's Concise style, and its replies have 54% fewer words for you to read. Every hidden test still passed. Measured on Claude Opus 5.5, with the method and every number in BENCHMARKS.md.
Use it to spend less of your Claude plan on each task and to get through each reply faster: the result, the test status and what went differently, without the paragraph of filler around them.
What you save
On everyday coding work Sharur costs 21% less than the built-in Concise style. That is the median of four medium tasks on Claude Opus 5.5: fixing failing tests, adding a CLI flag, reviewing three modules for bugs and tracing how an order total is computed (between 4% and 26% per task). Bigger jobs save less, and every result is in the table:
| Task | Tasks | Cost | Output tokens | Words to read |
|---|---|---|---|---|
| Short questions and small fixes | 4 | -37% | -35% | -61% |
| Long answers: an explanation, an audit | 2 | -21% | -23% | -38% |
| Medium coding tasks | 4 | -21% | -35% | -54% |
| Large features: returns, a move to integer cents | 2 | -10% | -13% | -47% |
| A huge task: an 11-item spec | 1 | -6% | 0% | -28% |
| Medium, large and huge, each size weighted equally | 7 | -12% | -16% | -43% |
Each task ran 3 times with each style, and each number is the median of the per-task changes. Cost is what the CLI reports, so it includes the style's own text, which is read on every request. Sharur was cheaper on all 13 tasks. Hidden tests checked the files every run left behind, and all of them passed with both styles. A grader found the same required facts in both styles' replies on the medium to huge tasks.
The words column is the time you spend reading. On a medium task Concise wrote a median of 129 to 380 words per task and Sharur 57 to 203, with the same facts in them.
Questions save the most, but few sessions are mostly questions, so 37% is not the number to expect. The bigger the task, the smaller the saving: a long session spends most of its tokens on code, tool calls and the context read again on every turn, and an output style changes none of that. JetBrains found the same for the caveman skill: 65% advertised, 8.5% on real agent tasks.
[!NOTE] Why the huge task shows 0% output tokens but 6% lower cost. There, almost everything the model writes is code: about 13,000 output tokens per run, of which the reply text is about 3%. Sharur's reply had 28% fewer words, about 120 tokens, which disappears in the run-to-run spread (9,700 to 17,400 output tokens with Concise). The cost fell for another reason: most of it is the context read again on every call, and Sharur made 10 calls in each run where Concise made 10 to 14. With one task and 3 runs, that 6% is inside the noise.
How this was measured, and how to rerun it, is in BENCHMARKS.md.
The whole style
Telegraphic: drop filler, keep grammar. <=2 lines, <=25 words. Bigger result: one short line per item, every item asked, no recap. No text before work is done. Say what was wrong, not a label for it. Exact: errors, destructive confirms. Fewest tool calls. Tool description fields <=3 words.
It is also smaller than the instructions Concise ships with, so every request carries about 330 fewer input tokens.
Before and after
Same prompt, same model (Opus 5.5), same small Python project: "Add a multiply function to calc.py with a test, and run the tests." The test file has a bug the prompt doesn't mention.
Concise, 89 words, 1,619 output tokens:
I added
multiplyand its test, but the test run fails because of a bug that was already insubtract.File "test_calc.py", line 4, in <module> assert subtract(5, 3) == 2 AssertionError
subtractreturnsa + binstead ofa - b(calc.py:6). The test file stops at the first failed check, so the newmultiplycheck never ran there. I ran it separately and it passes.I didn't change
subtractbecause you didn't ask me to. It's a one-character fix — want me to make it?
Sharur, 37 words, 1,012 output tokens:
Added
multiplytocalc.pyand a test, but the tests fail first atassert subtract(5, 3) == 2, so the multiply test never ran.subtractreturnsa + b; should I change it toa - b?
On a bigger result it writes one line per item. Sharur after "The tests fail. Find the causes and fix them." on a 9-file project:
All 19 tests pass now. The three causes:
shop/inventory.py:reserverefused an order for the last units because it checked>=instead of>.shop/pricing.pybulk_discount: the 10-unit check ran first, so 50+ units got 5% instead of 10%.shop/pricing.pyorder_totals: tax was charged before the coupon was taken off; it's now charged after.
Install
Paste this into Claude Code:
Install the Sharur output style for Claude Code.
1. Create the file ~/.claude/output-styles/sharur.md (create the folder if it is missing) with exactly this content:
---
name: Sharur
description: Telegraphic replies of 1-2 lines, one short line per item on bigger results, exact errors, as few tool calls as possible.
keep-coding-instructions: true
---
Telegraphic: drop filler, keep grammar. <=2 lines, <=25 words. Bigger result: one short line per item, every item asked, no recap. No text before work is done. Say what was wrong, not a label for it. Exact: errors, destructive confirms. Fewest tool calls. Tool description fields <=3 words.
2. In ~/.claude/settings.json set "outputStyle" to "Sharur". Keep every other key as it is. Create the file if it does not exist.
3. Tell me to start a new session, because the output style is read when a session starts.
Or do it by hand:
mkdir -p ~/.claude/output-styles
curl -fsSL https://raw.githubusercontent.com/ilien-dev/sharur/main/sharur.md -o ~/.claude/output-styles/sharur.md
Then pick Sharur in /config under Output style.
[!IMPORTANT] Start a new session after installing: Claude Code reads the output style when a session starts.
On Windows, ~ is %USERPROFILE%. To use it in one project only, put the file in .claude/output-styles/ inside that project. To go back, pick another style in /config and delete the file.
How it compares
The three most popular token savers for Claude Code have been through the same independent test by JetBrains: 80 to 86 real agent tasks, paired runs, cost taken from the bill.
| What it changes | What you install | Advertised | Measured by JetBrains | |
|---|---|---|---|---|
| caveman, 121k stars | How the agent talks, plus a proxy, model routing and a cloud service | A plugin, or an installer (Node.js 22) that adds a CLI and hooks | 65% fewer output tokens | 8.5% fewer output tokens; about 10% lower cost per task, erased in the totals by one outlier run |
| rtk, 87k stars | Shrinks command output before the model reads it | A Rust binary and a hook | 60 to 90% fewer tokens | Median cost 7.6% higher at low effort, no change at high effort |
| ponytail, 193k stars | Makes the agent write less code | A plugin with a hook that runs at session start | 54% less code, 20% lower cost | 15% less code, 10.3% lower cost |
| Sharur | How the agent talks | One Markdown file | 21% lower cost on medium tasks | Not tested by JetBrains yet |
What that means for you:
- Nothing runs on your machine. No binary, hook, proxy, account or telemetry; caveman's CLI sends usage statistics by default, by its own README. To remove Sharur, delete one file.
- The limit is a number the model can follow. caveman's README measures its own voice at 3% shorter than the plain instruction "Answer concisely." (10 tries). Sharur's replies on medium tasks had 54% fewer words than Concise, Claude Code's built-in short style.
- Every number comes from the bill and from hidden tests. The cost is what the CLI reports, including the style's own text, and the files each run left behind were tested.
- It works next to ponytail. ponytail cuts the code the agent writes and Sharur cuts what it says, so they don't overlap. The two have not been measured together.
Two things in fairness. Sharur only covers Claude Code, while caveman supports more than 30 agents. And Sharur's numbers come from its own 13-task benchmark, not from JetBrains' 86, so the last row is not a like-for-like comparison with the three above it.
Before you switch
[!WARNING]
- On big tasks the saving is small: 6 to 10% of the cost in these tests. The larger numbers come from questions and short tasks.
- This version of the style was measured on Opus 5.5 only, 3 runs per task. The earlier one-line version was also tested on Sonnet 5.5, Haiku 5.5 and Fable 5.2; this one has not been. Fable 5.2 runs made while tuning it looked similar, but those runs were used to choose the text, so they prove nothing.
- "Fewest tool calls" changes how the model works, not only how it talks. It never cost a hidden test here, but it is the first part to delete if you see the model skip a step it should have taken.
- The reports are written for someone who knows the task. If you need a full write-up of a change, ask for one or stay on Concise.
Why there are no hooks
Claude Code hooks can't rewrite a reply. A Stop hook can only block it and send the model back for another turn, which reads the whole context again. A PreToolUse hook can shorten a tool call's input, but only after the model has written, and been billed for, those tokens. In these runs the reply text was 10% of what the model wrote on medium to huge tasks, and tool description fields were under 1%, so a hook has little to save and the extra turn costs more than that. A proxy that compresses tool output, rtk, raised the median cost by 7.6% in JetBrains' tests. Sharur stays a text file.
Sharur is the mace of the Sumerian god Ninurta. The name means "smasher of thousands".
License
MIT
// faq
What is sharur?
Cut Claude Code token usage and cost with a one-line output style. Shorter replies than Concise on every Claude model: 29% fewer output tokens and 23% lower cost on Opus 5.5. Benchmarked.. It is open-source on GitHub.
Is sharur free to use?
sharur is open-source under the MIT license, so it is free to use.
What category does sharur belong to?
sharur is listed under devtools in the Claudeers registry of Claude-compatible tools.
// embed badge
[](https://claudeers.com/sharur)
// retro hit counter
[](https://claudeers.com/sharur)
// reviews
// guestbook
// related in Developer Tools
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Curs…
Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.
AI coding assistant skill (Claude Code, Codex, OpenCode, Cursor, Gemini CLI, and more). Turn any folder of code, SQL schemas, R scripts, shell scripts, docs,…
Use Garry Tan's exact Claude Code setup: 23 opinionated tools that serve as CEO, Designer, Eng Manager, Release Manager, Doc Engineer, and QA