lean-code 💰

English | 中文说明

Token-saving skills for Claude Code, Codex, Cursor & Windsurf.

caveman saves tokens by talking like caveman. lean-code saves tokens by being smarter about what gets sent.

Your output tokens are cheap. Your input tokens — the files you load, the history you carry — are where the real bill is.

---

What it does

4 skills. Each targets a different source of token waste:

Skill	What it fixes	Typical savings
/diff-only	AI rewrites full files for a 3-line fix	60–80% output tokens
/smart-read	AI reads 500-line files to find one function	40–90% input tokens
/trim	Conversation history balloons over long sessions	55–70% input tokens
/checkpoint	Context lost when switching tasks or sessions	Compresses 20k tokens → ~500

---

Benchmark

Real measurements. Claude claude-opus-4-5. 10 prompts per skill.

Skill	Output token savings	Total token savings
/diff-only	71% output	38% total
/smart-read	12% output	61% total
/trim	8% output	58% total
/checkpoint	45% output	28% total

Last run: 2026-05-22 · Model: claude-opus-4-5

Want to verify? Run python3 benchmarks/run.py with your own API key.

---

Install

One-liner (auto-detects all agents on your machine):

curl -fsSL https://raw.githubusercontent.com/EA-Studio-SHARK/lean-code/main/install.sh | bash

Per-agent:

Agent	Command
Claude Code	claude skills install EA-Studio-SHARK/lean-code
Codex	npx skills add EA-Studio-SHARK/lean-code -a codex
Cursor	npx skills add EA-Studio-SHARK/lean-code -a cursor
Windsurf	npx skills add EA-Studio-SHARK/lean-code -a windsurf

---

Usage

/diff-only      # all code changes shown as diffs from now on
/smart-read     # grep before reading files
/trim           # compress this session's history
/checkpoint     # save portable context snapshot

/diff-only in action

Without lean-code — AI rewrites the whole file:

// Returns 847 tokens of unchanged code + your 3-line fix
class UserService {
  private db: Database;
  constructor(db: Database) { this.db = db; }
  // ... 60 lines you didn't ask to see ...
  async getUser(id: string) {
    return this.db.users.findOne({ id }); // ← the actual fix
  }
}

With /diff-only — AI outputs only what changed:

// filepath: src/services/user.service.ts
- async getUser(id: number) {
+ async getUser(id: string) {
-   return this.db.users.findOne({ id: id });
+   return this.db.users.findOne({ id });
  }

Output: 28 tokens instead of 847. That's 97% less.

---

/smart-read in action

Without lean-code:

> read src/auth/middleware.ts  ← 340 lines, 2,380 tokens

With /smart-read:

> grep -n "authenticate" src/auth/middleware.ts
  → line 47
> read lines 40-65
  ← 26 lines, 182 tokens

Input: 182 tokens instead of 2,380. That's 92% less.

---

/trim in action

2-hour session: ~18,000 input tokens of conversation history.
After /trim: ~5,400 tokens. Saves ~70% on every subsequent message.

---

Cost impact

At Claude claude-opus-4-5 pricing ($3/MTok input, $15/MTok output):

Session type	Without lean-code	With lean-code	Monthly saving
4hr coding session/day	~$8.40/day	~$2.90/day	~$165/month
Background agent (Codex-style)	~$12/day	~$4.10/day	~$236/month

Numbers based on real benchmark runs. Your mileage varies with codebase size.

---

How it compares to caveman

caveman is brilliant. It saves output tokens by changing how Claude talks.

lean-code saves input tokens by changing what Claude reads.

Use both. They don't overlap.

# install caveman (output savings)
curl -fsSL https://raw.githubusercontent.com/JuliusBrussee/caveman/main/install.sh | bash

# install lean-code (input savings)
curl -fsSL https://raw.githubusercontent.com/EA-Studio-SHARK/lean-code/main/install.sh | bash

---

Run benchmarks yourself

git clone https://github.com/EA-Studio-SHARK/lean-code
cd lean-code
pip install anthropic
export ANTHROPIC_API_KEY=sk-ant-...
python3 benchmarks/run.py

Results are saved to benchmarks/results/ and README is auto-updated.

---

🔗 Related Projects

• awesome-mcp-zh — 🇨🇳 MCP 中文精选指南，国内可用的 MCP 服务器列表
• china-hot-mcp — 🔥 中国实时热搜 MCP 服务器，让 AI 实时获取中文热点

---

📖 中文说明

lean-code 是什么？

省 Token 神器 — 给 Claude Code、Cursor、Windsurf、Codex 装上 4 个技能，实测节省 40-80% 的 Token 消耗和费用。

你的输出 Token 其实不贵。真正烧钱的是输入 Token — 你加载的文件、携带的对话历史。

它解决什么问题？

技能	解决的问题	典型节省
/diff-only	AI 为了改 3 行代码，重写整个文件	60-80% 输出 Token
/smart-read	AI 读完 500 行文件只为找一个函数	40-90% 输入 Token
/trim	长对话的历史越滚越大	55-70% 输入 Token
/checkpoint	切换任务/会话时丢失上下文	2万 Token → ~500

实测数据

基于 Claude claude-opus-4-5，每个技能 10 条 prompt：

技能	输出节省	总节省
/diff-only	71%	38%
/smart-read	12%	61%
/trim	8%	58%
/checkpoint	45%	28%

省多少钱？

按 Claude claude-opus-4-5 定价（输入 $3/百万Token，输出 $15/百万Token）：

使用场景	不用 lean-code	用 lean-code	每月省
每天 4 小时编程	~$8.40/天	~$2.90/天	~$165/月
后台 Agent（Codex 模式）	~$12/天	~$4.10/天	~$236/月

一键安装

curl -fsSL https://raw.githubusercontent.com/EA-Studio-SHARK/lean-code/main/install.sh | bash

和 Caveman 的区别

Caveman 省输出 Token（改变 AI 说话方式）。
lean-code 省输入 Token（改变 AI 读取内容）。
两者不冲突，建议同时使用。

---

License

MIT