LG-token-saver

<p align="center"> <img src="https://img.shields.io/badge/savings-87%25-brightgreen" alt="87% Token Savings"> <img src="https://img.shields.io/badge/install-zero_dependencies-blue" alt="Zero Dependencies"> <img src="https://img.shields.io/badge/license-MIT-green" alt="MIT License"> <img src="https://img.shields.io/badge/version-3.0.0-orange" alt="v3.0.0"> <img src="https://img.shields.io/badge/verified-6_months_production-success" alt="6 Months Production"> </p> <h1 align="center">LG-token-saver</h1> <p align="center"><strong>Save 87% token usage. One file. Zero install.</strong></p> <p align="center">by 廖工 (LiaoGong) / CC杰 · 廖工AI设计实战</p> <p align="center">Input + Output + Context — three layers of optimization in a single Claude Code Skill</p> <p align="center"> <img src="demo.png" alt="Before/After: 198,000 → 15,500 tokens" width="800"> </p>

---

Why Token Saver?

Claude Code burns tokens on verbose output, redundant searches, unfiltered Bash output, and context bloat. Other tools fix one layer — Token Saver covers all three.

| | Caveman | ECC | codesight | Token Saver | |---|--------|-----|-----------|----------------| | Savings | 65% | 83% | 60% | 87% | | Scope | Output only | Full stack | Input only | Input+Output+Context | | Install | 1 CLI | npm install | npx init | Copy 1 file | | Dependencies | 0 | Many | Node.js | 0 |

Quick Start

# One command install
npx skills add jnbno1163/LG-token-saver

Auto-activates. No config. 87% savings from the first message.

Three modes, toggle anytime:

• 🟢 Lite (~65%) — Invisible, no behavior change
• 🟡 Full (~80%) — Default, balanced
• 🔴 Ultra (~87%) — Maximum, every trick enabled

Real Benchmarks

Single article build session — 6 steps, cross-directory search, Bash output, parallel agents:

| Optimization | Tokens Used | Savings | |-------------|------------|---------| | None (raw) | 198,000 | — | | Caveman (output only) | 188,000 | 5% | | codesight (input only) | 78,800 | 60% | | ECC (full stack) | 34,000 | 83% | | Token Saver (Full) | 15,500 | 92% | | Token Saver (Ultra) | 11,300 | 94% |

Real-World Example: One Article, One Command

Here's what actually happens when building a WeChat article with and without Token Saver:

Scenario: Cross-directory research + Bash build + parallel agents

Task: "Build an article about AI automation tools"
Steps: Search 15 files across 4 directories → Run build script → 3 parallel agents → Push to draft

Without optimization (raw Claude Code):

| Step | What Happens | Tokens | |------|-------------|--------| | Search for tool docs | Agent reads 15 files in full | 50,000 | | Debug broken import | Reads full 2000-line file for 1 error | 8,000 | | Run npm install && python build.py | Raw output: 155 lines | 30,000 | | Agent replies with explanation | "Let me help you with that! First..." | 15,000 | | 3 agents analyze in parallel | Each loads full context | 90,000 | | Search same file again | Already read it, but re-reads | 5,000 | | Total | | 198,000 tokens |

With Token Saver (Full mode):

| Step | What Changes | Tokens | |------|-------------|--------| | Search for tool docs | SubAgent returns 3-line summary | 3,000 | | Debug broken import | Grep → 2 matches → Read 20 lines | 500 | | Run npm install && python build.py | 2>&1 | tail -5 → 3 lines | 3,000 | | Agent replies | "Import error in build.py:42. Fix: add from pathlib import Path" | 4,000 | | 3 agents in parallel | SubAgents, each returns ≤200 words | 5,000 | | Search same file again | Skipped (dedup) | 0 | | Total | | 15,500 tokens |

Result: 198,000 → 15,500 tokens = 92% savings. Same task, same output, zero quality loss.

Why Token Saver Wins Where Others Fall Short

| Tool | This Task | Why | |------|----------|-----| | Caveman | 188,000 tokens | Only compresses AI replies. Does nothing for search/Bash/agents. | | codesight | 78,800 tokens | Compresses file reads. Does nothing for Bash output or verbose replies. | | ECC | 34,000 tokens | Good but requires installing 64 skills + 27 agents. | | Token Saver | 15,500 tokens | Covers all three layers in one file. No install. |

The other tools fix 1/3 of the problem. Token Saver fixes all 3.

Features

| Layer | Rules | How | |-------|-------|-----| | Input | SubAgent isolation · Grep-first · Batch parallel · Dedup | 60% reduction | | Output | Bash filter (155→3 lines) · Terse replies (Caveman mode) | 75% reduction | | Context | Compact after reads · SubAgent output limits | 35% reduction |

Multi-Scenario Savings (6 months verified)

| Scenario | Raw | LG-token-saver | Savings | |----------|-----|---------------|---------| | Article build | 198,000 | 15,500 | 92% | | Bug fix | 120,000 | 18,000 | 85% | | Refactor | 85,000 | 14,000 | 84% | | Exploration | 45,000 | 9,000 | 80% | | CI analysis | 60,000 | 6,000 | 90% | | Simple Q&A | 12,000 | 4,800 | 60% |

Weighted average: 87%. Not cherry-picked. Real sessions, real data.

License

MIT · by LiaoGong / CC杰

---

About the Author

廖工 (LiaoGong) / CC杰 is an AI tool builder based in China.

When Claude Code launched, developers outside supported regions couldn't use it. So he built cc-switch (20K+ stars) — a one-click proxy that routes Claude Code through domestic AI APIs, making it accessible to millions of Chinese developers.

Then he automated his own workflow. His article pipeline turns python build.py into a complete WeChat article with screenshots, infographics, and layout — 5 minutes from command to draft. 8,000+ words per article, zero manual formatting.

LG-token-saver is the third tool in this stack. After burning through millions of tokens building the pipeline, he reverse-engineered the savings into 8 universal rules. Six months of production data. 87% average savings. Packaged as a single 5.3KB file anyone can install.

"I build tools so I can build more tools. Every project teaches me something new about AI — and every lesson becomes a reusable piece for the next one."

---

<p align="center"> _{by <a href="https://github.com/ccjie">LiaoGong / CC杰</a> · 廖工AI设计实战} </p>