Sui (粋) - A Programming Language for LLMs

Zero syntax errors. Zero typos. LLMs generate code that just works.

日本語版 README

Try it Now

No installation required - try Sui directly in your browser:

• Sui Playground - Write and run Sui code instantly
• Counter Demo - See Sui + WebAssembly in action

Overview

Sui (粋) is a programming language named after the Japanese aesthetic concept meaning "refined" and "elimination of excess." It is designed so that LLMs can generate 100% accurate code - not through hope, but through structural guarantees.

Why Sui?

Current LLM code generation problems:

• Bracket mismatches if (x { }
• Variable typos coutn vs count
• Indentation errors
• Complex nested expressions

Sui makes these errors structurally impossible:

Problem	Sui's Solution
Bracket mismatch	Only {} for functions, no nesting
Variable typos	Sequential numbers only (v0, v1, g0) - can't misspell
Indentation errors	Line-based, indentation is ignored
Complex nesting	One instruction per line, decompose to temps

Design Principles

1. Zero Syntax Error Rate - Structurally impossible to make syntax errors
2. Zero Typo Rate - Variables are numbers, not names
3. Line Independence - Each line is completely self-contained
4. Pure Logic Language - Computation only; UI delegated to any framework
5. Future-Proof Efficiency - As LLMs learn Sui, token efficiency will surpass traditional languages

Installation

# PyPI (basic)
pip install sui-lang

# PyPI with WebAssembly support
pip install sui-lang[wasm]

# Homebrew (macOS/Linux)
brew tap TakatoHonda/sui
brew install sui-lang

# From source
git clone https://github.com/TakatoHonda/sui-lang.git
cd sui-lang

Quick Start

Interactive Mode (REPL)

# Start REPL
sui

# Example session
>>> = v0 10
>>> + v1 v0 5
>>> . v1
>>>
15
>>> .exit

Commands: .exit / .quit (exit), .reset (reset state)

Interpreter

# Run file
sui examples/fibonacci.sui

# Run with arguments
sui examples/fib_args.sui 15

# Validate
sui --validate examples/fibonacci.sui

# Show help
sui --help

Transpiler (Sui → Python)

# Show converted code
sui2py examples/fibonacci.sui

# Output to file
sui2py examples/fibonacci.sui -o fib.py

# Convert and execute
sui2py examples/fib_args.sui --run 15

Transpiler (Python → Sui) for humans

# Show converted code
py2sui your_code.py

# Output to file
py2sui your_code.py -o output.sui

WebAssembly

# Compile to WebAssembly binary (requires: brew install wabt)
sui2wasm examples/fibonacci.sui -o fib.wasm

# Execute directly via WebAssembly (requires: pip install sui-lang[wasm])
suiwasm examples/fibonacci.sui

Browser UI

Sui is a pure logic language. UI can be implemented with any framework (React, Vue, Hono.js, vanilla JS, etc).

Sui compiles to Wasm with exports:

• main() - Initialization
• f0(), f1(), ... - Functions (callable from JS)
• g0, g1, ... - Global variables (read/write via .value)

// Any framework works
const wasm = await WebAssembly.instantiate(wasmBytes, { env: { print_i32: console.log }});
button.onclick = () => { wasm.exports.f0(); display.textContent = wasm.exports.g0.value; };

Running without Installation (from source)

# Using python directly
python sui.py examples/fibonacci.sui
python sui2py.py examples/fibonacci.sui
python py2sui.py your_code.py

Syntax

Instructions

Instr	Format	Description
=	= var value	Assignment
+	+ result a b	Addition
-	- result a b	Subtraction
*	* result a b	Multiplication
/	/ result a b	Division
%	% result a b	Modulo
<	< result a b	Less than (0/1)
>	> result a b	Greater than (0/1)
~	~ result a b	Equality (0/1)
!	! result a	NOT
&	& result a b	AND
|	| result a b	OR
?	? cond label	Conditional jump
@	@ label	Unconditional jump
:	: label	Label definition
#	# id argc {	Function definition start
}	}	Function definition end
$	$ result func args...	Function call
^	^ value	Return
[	[ var size	Array create
]	] result arr idx	Array read
{	{ arr idx value	Array write
.	. value	Output
,	, var	Input

Variables

Format	Meaning
v0, v1, ...	Local variables
g0, g1, ...	Global variables
a0, a1, ...	Function arguments
c0	argc (command-line argument count)
c1, c2, ...	argv (command-line arguments, read-only)

Examples

Fibonacci

# 0 1 {
< v0 a0 2
! v1 v0
? v1 1
^ a0
: 1
- v2 a0 1
$ v3 0 v2
- v4 a0 2
$ v5 0 v4
+ v6 v3 v5
^ v6
}
= g0 10
$ g1 0 g0
. g1

Output: 55

FizzBuzz

= v0 1
: 0
> v1 v0 100
? v1 9
% v2 v0 15
~ v3 v2 0
? v3 1
% v4 v0 3
~ v5 v4 0
? v5 2
% v6 v0 5
~ v7 v6 0
? v7 3
. v0
@ 4
: 1
. "FizzBuzz"
@ 4
: 2
. "Fizz"
@ 4
: 3
. "Buzz"
@ 4
: 4
+ v0 v0 1
@ 0
: 9

File Structure

sui/
├── README.md           # This file (English)
├── README_ja.md        # Japanese README
├── LICENSE             # MIT License
├── sui.py              # Interpreter
├── sui2py.py           # Sui → Python transpiler
├── sui2wasm.py         # Sui → WebAssembly binary compiler
├── suiwasm.py          # WebAssembly runtime (execute via wasmtime)
├── py2sui.py           # Python → Sui transpiler (for humans)
├── examples/
│   ├── fibonacci.sui
│   ├── fib_args.sui
│   ├── fizzbuzz.sui
│   ├── list_sum.sui
│   ├── args_demo.sui
│   └── counter_app/      # Full app example (Sui + Wasm + HTML)
└── prompts/
    ├── system_prompt_en.md  # LLM system prompt (English)
    ├── system_prompt_ja.md  # LLM system prompt (Japanese)
    └── examples.md          # Application examples (Sui + UI)

LLM Integration

Sui is designed for LLM code generation. Use the prompts in prompts/ directory:

1. Copy the system prompt from prompts/system_prompt_en.md
2. Paste it into ChatGPT / Claude / Gemini / etc.
3. Ask to generate Sui code for your task
4. Run with sui your_code.sui

See prompts/examples.md for prompt templates and expected outputs.

Name Origin

Sui (粋) - A Japanese word meaning "refined," "sophisticated," or "the essence." It represents the aesthetic of eliminating excess and keeping only what is essential.

Token Efficiency: Now vs Future

Current state (LLMs don't know Sui yet):

Language	Fibonacci	Counter
Sui	79 tokens	44 tokens
Python	30 tokens	30 tokens

Future state (after LLMs learn Sui):

• v0, g0 → 1 token each (currently 2)
• Patterns like + g0 g0 1 → compressed
• Estimated: 40-50% reduction

But token count isn't the point. The real value:

• 0% syntax error rate (vs ~5-10% for Python/JS)
• 0% typo rate (variables can't be misspelled)
• 100% parseable output (every line is valid or clearly invalid)

vs Other Languages

vs Python/JavaScript

Aspect	Python/JS	Sui
Syntax errors	Common	Impossible
Variable typos	Common	Impossible
Bracket matching	Error-prone	Trivial
Token efficiency (now)	Better	Worse
Token efficiency (future)	Same	Better

vs Assembly

Aspect	Assembly	Sui
Instructions	Hundreds	~20
Registers	8-32	Unlimited
Learning curve	Steep	Minimal

Roadmap

• Interpreter (Python)
• Transpiler (Sui → Python)
• Transpiler (Python → Sui, for humans)
• Interactive mode (REPL)
• WebAssembly output (WAT + runtime)
• Package manager (hash-based IDs) (#9)
• Standard packages: sui-math, sui-crypto (#8)
• Type annotations (optional)
• LLVM IR output

Future: Standard Packages

Mathematical and utility functions will be provided as standard packages (not built-in):

; sui-math package (48-bit hash ID)
X 182947362847591 0 v2 v0 v1   ; matmul(v0, v1) → v2
X 182947362847591 1 v3 v0      ; transpose(v0) → v3
X 182947362847591 10 v4 v0     ; mean(v0) → v4
X 182947362847591 11 v5 v0     ; std(v0) → v5

; sui-crypto package
X 56284719384756 0 v6 v0       ; sha256(v0) → v6

Design principles maintained:

• No identifiers (package/function IDs are numeric)
• Consistent with package manager design (#9)
• Core language stays minimal

See Issue #8 (sui-math) and Issue #9 (package manager) for details.

License

MIT License