⚡ ZipAI

⭐ If you find ZipAI helpful for reducing LLM API token costs, please consider giving it a star on GitHub! ⭐

• 🧹 Zero Filler — strips conversational padding, enforces telegraphic grammar.
• 💾 Prompt Caching — static-first ordering preserves prefix cache hits (up to 90%).
• ✂️ Input Pruning — log compression, AST skeletal inspection, JSON/YAML minification.
• 🔬 Surgical Output — localized diffs only, no full-file reprints.
• 🧠 Reasoning Budget — adaptive Chain-of-Thought depth based on task complexity.
• 📐 Schema Enforcement — JSON Schema replaces verbose format instructions.
• 🗜️ Semantic Compression — lossless context reduction (50–90% savings).
• 🧭 ARM-style Reasoning — task-aware reasoning mode selection.
• ⚡ KV Cache Optimization — structured inter-tool communication, dynamic eviction.
• 📊 TAAC — domain-specific compression ratios (code vs math vs NL).

Install the rules into your AI agent. Pick the path for your tool:

# Antigravity IDE (Gemini Agent Skill)
mkdir -p ~/.gemini/config/skills/zipai
cp SKILL.md ~/.gemini/config/skills/zipai/SKILL.md

# Claude Code CLI
cp CLAUDE.md /path/to/your/project/CLAUDE.md

Restart your agent — it now applies ZipAI rules automatically. ✅

mkdir -p ~/.gemini/config/skills/zipai
cp SKILL.md ~/.gemini/config/skills/zipai/SKILL.md

The IDE lists zipai-optimizer as an available skill and applies the rules automatically.

cp CLAUDE.md /path/to/your/project/CLAUDE.md

Claude Code reads this file at startup and applies the rules for the workspace session.

mkdir -p /path/to/your/project/.agents/rules/
cp .agents/rules/zipai.md /path/to/your/project/.agents/rules/zipai.md

cp SKILL.md /path/to/your/project/GEMINI.md

These rules are framework-agnostic markdown. Load SKILL.md into any system-instructions template:

• Hermes Agent: append to the agent's system instructions profile or config.yaml.
• Custom loops: reference SKILL.md directly in system instructions.

ZipAI is a ruleset, not a runtime — once installed, your agent self-applies the optimizations. Verify behavior by prompting normally and observing denser, cache-friendly output:

# In your agent session
"Summarize this 50-line log and propose a fix."   # expect ~225 tokens, not ~4,500

Tip: CLAUDE.md and .agents/rules/zipai.md are mirrors of SKILL.md. Update SKILL.md first, then propagate.

ZipAI compresses tokens at every stage of the agent loop:

flowchart LR
  A[Raw Prompt / Output] --> B[Zero Filler]
  B --> C[Prompt Caching]
  C --> D[Input Pruning]
  D --> E[Semantic Compression]
  E --> F[Ultra-Dense Tokens]
  F --> G[LLM Context + Cache]

ZipAI/
├── SKILL.md                    # Canonical rules (v15.0) — single source of truth
├── CLAUDE.md                   # Mirror for Claude Code auto-detection
├── .agents/rules/zipai.md     # Mirror for VS Code agent extensions
├── README.md                   # This file
├── LICENSE                     # MIT License
├── .pre-commit-config.yaml    # Git hooks
├── .gitattributes              # Git merge strategies
└── .github/
    ├── scripts/ai-pr-reviewer.cjs
    └── workflows/
        ├── ai-pr-reviewer.yml
        └── auto-resolve-pr-conflicts.yml

• Brainstorming: disable during creative / open-ended design phases.
• Grep Blindness: key context may fall outside filter boundaries.
• Overshadowing: aggressive pruning may drop micro-variables in long sessions.
• Math Fragility: numerical reasoning chains resist aggressive compression.

If ZipAI helped optimize your agent workflows or slashed your API bill, leave a star ⭐ to help other developers discover it!

ZipAI ships markdown rules only — no executable code. See SECURITY.md for the vulnerability-reporting policy.

This project is licensed under the MIT License.