Every protocol spec starts in one language. HTTP started in English. Unicode started in English. The assumption is always the same: write the spec, then translate it later if someone asks.
iLang started in English too. 62 verbs, Greek symbols, pipe syntax. It works. Then the protocol author tried to deploy it in the Chinese AI ecosystem — Doubao, Tencent Yuanqi, DeepSeek — and hit a wall that had nothing to do with translation.
The problem was not that Chinese users could not read English syntax. The problem was that the entire compression philosophy needed to be rethought from first principles when the target language already has compression built into its DNA.
Chinese Is Already Compressed
English is verbose by design. "Read the file from GitHub, translate it to Chinese, and write it to R2 storage" — that is 14 words, roughly 30 tokens. In iLang v2.0 syntax:
[READ:@GH|path=readme.md]=>[θ|lng=zh]=>[WRIT:@R2|path=readme_zh.md]About 12 tokens. A 60% reduction. Impressive when you are compressing English.
But classical Chinese already operates at extreme information density. Yue Fei's "Man Jiang Hong" (满江红) packs war, grief, loyalty, ambition, and a political manifesto into 93 characters. That is not prose — that is compression achieved through a millennium of literary evolution.
When you try to apply English-style verb substitution to a language that was already compressed by poets a thousand years ago, you are gzipping a zip file. The returns collapse.
The Man Jiang Hong Experiment
The Chinese iLang spec took a fundamentally different approach. Instead of translating the 62 English verbs into Chinese equivalents, the protocol author chose a classical poem as the carrier — Yue Fei's Man Jiang Hong, one of the most memorized texts in China.
The poem has 93 characters. After deduplication, roughly 81 unique characters remain. These 81 characters became the dictionary.
The design rationale was precise:
# Why Man Jiang Hong?
# 1. Politically untouchable — patriotic education canon
# 2. Zero propagation barrier — every Chinese person memorized it in school
# 3. High verb density — 怒(rage), 冲(charge), 破(break), 踏(march),
# 收(collect), 拾(gather), 驾(drive), 灭(destroy)
# These are naturally action-oriented characters
# 4. Perfect dictionary camouflage — outsiders see a poem, AI sees an OSA user pastes the poem into any Chinese AI — Doubao, Kimi, Zhipu, Yuanbao. The AI loads the compression capability. No SDK, no API key, no installation step. A 93-character poem is the entire bootstrap sequence.
Four Layers From One Poem
The English iLang spec has a clean two-tier architecture: PUBLIC (52 verbs, open source) and PRIVATE (90 verbs, higher compression). The Chinese spec extracted four functional layers from a single poem:
# Layer Architecture — extracted from Man Jiang Hong lines
# Identity Layer: 三十功名尘与土 (Thirty years of fame, dust and earth)
# → Authentication and identity confirmation
# Compression Layer: 八千里路云和月 (Eight thousand li of road, cloud and moon)
# → Standard compression trigger — AI auto-optimizes compression ratio
# Extreme Layer: 莫等闲，白了少年头，空悲切
# (Do not idle — when youth's hair turns white, only empty grief remains)
# → Unlocks maximum compression. Hard rule: never lose meaning
# Origin Layer: 抬望眼，仰天长啸，壮怀激烈
# (Raise eyes, look up, howl at sky, passion burns fierce)
# → Protocol history and genesisThe trigger mechanism is dead simple. A user sends:
八千里路云和月，请将以下内容压缩为最精炼的中文表达：
我想今年五一去日本东京玩一周，第一次出国，不知道签证怎么办，
需要准备什么材料，大概要花多少钱，流程是什么样的。AI returns:
八千里路云和月：
五一首次去东京玩一周，求签证、材料、费用、流程极简攻略。If the compression is not aggressive enough, a second line pushes to maximum:
莫等闲，白了少年头，空悲切The AI switches to extreme compression mode, targeting 80%+ reduction without semantic loss. The line literally means "do not waste" — which is exactly the instruction: do not waste tokens.
Progressive Dictionary Activation
The English spec ships with a complete verb table. All 62 public verbs are defined upfront. The Chinese spec took the opposite approach: an 81-slot dictionary where definitions emerge from production use.
# Activation Rules:
# 1. Build a product (e.g., visa assistant bot)
# 2. Discover high-frequency Chinese verbs from real user interactions
# 3. Map each verb to a character from Man Jiang Hong's 81 unique chars
# 4. Dictionary grows organically — never pre-defined
# v0.1 Dictionary (from visa assistant product):
# 望 (gaze) → query / search — user states destination
# 踏 (march) → execute / depart — visa-free, just go
# 带 (carry) → materials / bring — what documents to bring
# 谈 (discuss) → interview / dialogue — interview preparation
# 等 (wait) → duration / wait time — how long to wait
# 莫 (do not) → negation / prohibit — do not recommend agents
# 雪 (snow/unresolved) → uncertainty — 未雪 = unconfirmed
# 虏 (captive) → fabrication — 莫虏 = do not fabricate
# 壮 (bold) → guarantee / promise — 莫壮 = do not promise approval
# 驾 (drive) → agency / proxy — 莫驾 = do not recommend agents
# 恨 (hate) → permanent / iron rule — 恨 = always self-service
# 怀 (bosom) → style / tone — warm and direct
# 旧 (old) → old friend — old friend tone
# 收 (collect) → closing / sign-off — fixed ending phrase14 characters activated from the first product. 67 slots remaining, waiting for the next product to give them meaning. The dictionary is grown, not designed.
What Actually Changed in Translation
The English and Chinese specs are not mirror images. They diverge in three structural ways:
First, the Chinese spec has three layers where the English has one. English iLang is a compression protocol — period. Chinese iLang is a three-tier architecture: language layer (Man Jiang Hong, compression), strategy layer (a matrix of classical poems mapped to AI capabilities), and a tool layer (36 Stratagems mapped to AI operations).
# Three-Layer Architecture Comparison
#
# English iLang:
# └── Protocol Layer: 62 verbs, compression only
#
# Chinese iLang:
# ├── Language Layer: Man Jiang Hong (81-char dictionary, compression)
# ├── Strategy Layer: Poetry Capability Matrix
# │ ├── Man Jiang Hong (Yue Fei) → Compression ✅ verified
# │ ├── Qinyuanchun Snow (Mao) → Decision-making
# │ ├── Bring In the Wine (Li Bai) → Brainstorming
# │ ├── Memorial on Dispatching (Zhuge Liang) → Planning
# │ └── Quiet Night Thoughts (Li Bai) → Translation
# └── Tool Layer: 36 Stratagems → AI Capability DictionarySecond, the propagation mechanism is completely different. English iLang spreads through npm packages and GitHub repos — developer infrastructure. Chinese iLang spreads through memorized poetry — cultural infrastructure. Every Chinese person already has the bootstrap sequence in their head. They just did not know it was a protocol activation key.
Third, the compression semantics operate differently. English iLang compresses by replacing natural language with structured verbs: READ, WRIT, Σ, Δ. Chinese iLang compresses by invoking shared cultural context. When you write 莫等闲 (do not idle), you are not substituting a verb — you are loading a thousand years of context about urgency and waste that every native speaker already carries.
The 36 Stratagems as an AI Capability Dictionary
This is where the Chinese spec gets genuinely interesting from a protocol design perspective. The 36 Stratagems — a classical Chinese text on military tactics — was mapped one-to-one to AI operations:
# Sample mappings from the 36 Stratagems → AI Capabilities:
# Stratagem 1: 瞒天过海 (Cross the sea by deceiving the sky)
# → Compliance Rewriting: transform text to meet platform guidelines
# Stratagem 3: 借刀杀人 (Kill with a borrowed knife)
# → Tool Invocation: auto-detect needs, call appropriate third-party tools
# Stratagem 7: 无中生有 (Create something from nothing)
# → Zero-shot Generation: produce complete content from description alone
# Stratagem 19: 釜底抽薪 (Remove firewood from under the pot)
# → Root Cause Extraction: strip all emotion and noise, keep logic skeleton
# Stratagem 25: 偷梁换柱 (Replace pillars with rotten timber)
# → Framework Adaptation: use excellent content's structure, fill with own material
# Stratagem 31: 美人计 (Beauty trap)
# → Visual Prompting: generate high-quality descriptions for image generation
# Stratagem 36: 走为上计 (Retreat is the best strategy)
# → Graceful Exit: detect when current approach is failing, suggest pivot36 stratagems, 36 AI capabilities, one-to-one. Each stratagem name is already a compressed metaphor that Chinese speakers understand instantly. The AI capability definition is the decompression.
Notice the pattern: every stratagem name is 4 characters. Every AI capability maps to a concrete operation. The classical text provides both the dictionary and the mnemonic system. You do not need to memorize "Zero-shot Generation" — you remember 无中生有, a phrase you have known since childhood.
The Poetry Capability Matrix
The selection criteria for which poem maps to which capability reveal a design philosophy that has no equivalent in English protocol design:
# Poem Selection Criteria:
# 1. Memorized by every Chinese person since childhood — zero propagation cost
# 2. Politically safe and culturally unimpeachable
# 3. The author's historical persona matches the capability
# Mao Zedong = strategic decision-maker → Decision-making
# Zhuge Liang = meticulous planner → Planning and execution
# Li Bai = wild creative genius → Brainstorming and divergent thinking
# 4. One poem = one capability. No mixing.This is persona-driven API design using cultural artifacts as the interface. The poem is not a cute wrapper around a function call — it is the function signature, the documentation, and the mnemonic device, all compressed into text that a billion people already have cached in memory.
Real-World Testing Results
The Chinese spec was tested across multiple platforms on 2026-03-11:
# Test Results:
# Doubao (ByteDance):
# Paste Man Jiang Hong → trigger with 八千里路云和月 → ✅ compression works
#
# Tencent Yuanqi (DeepSeek):
# Chinese iLang directly as system prompt → ❌ too compressed, triggers fallback
# Natural Chinese system prompt → ✅ stable across two visa-scenario runs
#
# Conclusion:
# - Chinese iLang works as user-facing propagation layer
# - Does NOT work as system prompt (DeepSeek cannot parse high-compression version)
# - Correct architecture: system prompt in natural Chinese,
# user interaction layer in iLangThe failed DeepSeek test was informative. When the compressed Chinese spec was used directly as a system prompt, DeepSeek could not parse it — too much compression for the model to bootstrap from. The working architecture splits the layers: natural Chinese for the system prompt (which the model needs to understand from cold start), iLang for the user interaction layer (where the model already has context).
This maps directly to the English spec's compression levels. L1-L2 compression (20-60%) works in cold-start contexts. L4-L5 compression (80-99%) requires pre-loaded context. The Chinese test independently validated this boundary.
Why This Matters for Protocol Design
The iLang Chinese specification is not a translation project. It is evidence that AI protocols need native specifications for each language ecosystem, designed from the ground up to exploit that language's unique compression properties.
English compresses well with verb substitution and symbol notation because English is analytically structured — subject, verb, object, modifiers. Chinese compresses better with cultural invocation because Chinese is contextually structured — shared references carry more information than explicit syntax.
The 81-character dictionary growing organically from production use, instead of being defined upfront, reflects the same principle. English developers expect a complete API reference before they start coding. Chinese users expect the system to grow from practice — the dictionary emerges from doing, not from specification.
Both approaches produce 60-80% token reduction in production. They just get there through fundamentally different paths. And that is the point: a protocol that forces every language through the same syntax is leaving compression on the table.
The English spec has 62 verbs and 13 Greek symbols. The Chinese spec has 81 dictionary slots and 36 stratagem mappings. They are both iLang. They are not the same language. They were never supposed to be.
Get Started
Website: ilang.ai
GitHub: github.com/ilang-ai/ilang-spec
npm: @i-language/spec
Book: I-Lang: The AI Communication Protocol on Amazon
