FEW-AI-SERIAL: A Domain-Agnostic Semantic Compression Protocol for LLM Prompt Injection and IoT Data Encoding

We present FEW-AI-SERIAL, a domain-agnostic semantic compression protocol that reduces structured data payloads by 68–92% while maintaining full human readability and native interpretability by Large Language Models (LLMs). Unlike binary serialization formats (Protocol Buffers, Avro, CBOR) that achieve comparable compression but produce opaque byte streams, FEW-AI-SERIAL uses positional mnemonic keys (2–3 uppercase characters) and domain-native value notation to produce compact text that both humans and LLMs can read without external schema or decoder. We validate the protocol across three production domains: (1) LLM few-shot prompt injection for poker decision-making (86% token reduction, 100% extraction accuracy), (2) IoT sensor data encoding on ESP32 microcontrollers (92% byte reduction), and (3) Model Context Protocol (MCP) agent-to-agent communication (88% message size reduction). Empirical benchmarks using Claude Haiku demonstrate identical extraction accuracy (100%) with 41% fewer bytes and 31% lower latency compared to JSON.