sniffable-python（sniffable Python 测试）

Sniffable Python

"Structure code for the first 50 lines. That's all the LLM needs."

Don't invent new syntax — structure existing syntax for optimal LLM comprehension.

*If your code smells good, the LLM can sniff it from the first whiff.*

*Steve Jobs wanted pixels you could lick. We want code you can sniff.*

The Problem

LLMs can read Python. But:

• Large files overwhelm context
• Implementation details bury the API
• Comments scattered or missing
• No clear entry point for understanding

Solution: Structure Python so the first ~50 lines contain everything needed to understand and use the tool. Make your code's bouquet apparent from the opening notes.

The Pattern

Canonical Structure

#!/usr/bin/env python3
"""skill-name: Brief description of what the script does.

This docstring becomes --help output AND is immediately visible to the LLM.
It should contain:
- Purpose (one sentence)
- Usage examples
- Key behaviors

Usage:
    python script.py command [options]

Examples:
    python script.py move north --quiet
    python script.py examine sword --verbose
"""

import argparse
from pathlib import Path
import yaml

# Configuration
DEFAULT_ROOM = "start"
VALID_DIRECTIONS = ["north", "south", "east", "west"]
MAX_INVENTORY = 10

def main():
    """Main entry point — CLI structure only, no implementation."""
    parser = argparse.ArgumentParser(
        description=__doc__.split('\n')[0],  # First line of module docstring
        formatter_class=argparse.RawDescriptionHelpFormatter
    )
    subparsers = parser.add_subparsers(dest="command", required=True)

    # move command
    move_parser = subparsers.add_parser("move", help="Move in a direction")
    move_parser.add_argument("direction", choices=VALID_DIRECTIONS,
                             help="Direction to move")
    move_parser.add_argument("--quiet", "-q", action="store_true",
                             help="Suppress output")

    # examine command
    examine_parser = subparsers.add_parser("examine", help="Look at something")
    examine_parser.add_argument("target", help="What to examine")
    examine_parser.add_argument("--verbose", "-v", action="store_true",
                                help="Show details")

    # status command
    status_parser = subparsers.add_parser("status", help="Show game state")
    status_parser.add_argument("--verbose", "-v", action="store_true",
                               help="Show details")

    args = parser.parse_args()
    _dispatch(args)

# Implementation (LLM only reads past here if modifying behavior)

def _dispatch(args):
    """Route to appropriate handler."""
    if args.command == "move":
        _do_move(args.direction, args.quiet)
    elif args.command == "examine":
        _do_examine(args.target, args.verbose)
    elif args.command == "status":
        _show_status(args.verbose)

def _do_move(direction: str, quiet: bool) -> None:
    """Internal: Execute movement."""
    ...

def _do_examine(target: str, verbose: bool) -> None:
    """Internal: Examine an object."""
    ...

def _show_status(verbose: bool) -> None:
    """Internal: Display status."""
    ...

if __name__ == "__main__":
    main()

The Zones

Why This Works

Dual-Audience Design

# One file, two audiences, same source of truth
dual_audience:
  source: "Python file (sniffable)"
  audiences:
    human:
      entry_point: "--help"
      sees: "CLI usage, arguments, examples"
    llm:
      entry_point: "first 50 lines"
      sees: "docstring, imports, constants, CLI tree"
  convergence: "Same source of truth — no duplication, no drift"

DRY Principle

The CLI decorators ARE the documentation:

@click.option('--verbose', '-v', is_flag=True, help='Show details')

This single line defines:

• The flag name (--verbose)
• The short form (-v)
• The type (boolean flag)
• The help text
• The Python parameter name and type

No duplication. No drift. One source.

Why Syntax Compression Fails

Some argue for compressing syntax to save tokens — dense symbols, sigils, minimal keywords. This is the wrong optimization.

Perl is the cautionary tale. Its obsessive fetish for compact syntax, sigils, punctuation, and performative TMTOWTDI one-liners — to the point of looking like line noise — is exactly why it's no longer relevant for LLM comprehension and generation.

The LLM already knows Python cold. Teaching it a novel syntax costs millions of tokens per context window (docs, examples, corrections) — and that cost is paid every single prompt, because prompting is not training. LLMs have no memory. Sniffable Python costs zero tokens to teach — the LLM just reads code it understands deeply from massive pre-training.

The Symbol Collision Problem

Short symbols cause semantic collisions. When @ appears, it activates Python decorators, shell patterns, email addresses, and more — simultaneously. The model has to do disambiguation work that wouldn't exist with unambiguous keywords. That disambiguation costs compute and introduces error modes.

Comprehension Fidelity > Token Count

You're not optimizing for token count. You're optimizing for how well the LLM understands your code.

Dense punctuation is anti-optimized for how transformers tokenize and reason. Verbose, well-structured code with semantic comments actually compresses better in the LLM's internal representations.

Comments as YAML Jazz

Apply YAML-JAZZ principles to Python comments:

# CONSTANTS
TIMEOUT = 30        # generous — API is flaky on Mondays
MAX_RETRIES = 3     # based on observed failure patterns in prod
BATCH_SIZE = 100    # memory-safe for 8GB instances

# TODO: Add circuit breaker after next major outage
# FIXME: Rate limiting not yet implemented
# NOTE: This duplicates logic in api_client.py — consider extracting

These comments ARE data. The LLM reads them. Acts on them. Uses them to understand *why*, not just *what*.

Comments: Substance Over Decoration

Don't waste tokens on decorative separators:

# BAD: Decorative separators (long = or - lines add no meaning)
# GOOD: Plain and functional
# Imports
import argparse

Don't trim Christmas trees:

Avoid ASCII art or box-drawing in comments. Use plain section comments instead:

# Configuration
TIMEOUT = 30  # generous — API flaky on Mondays

Do include useful semantic comments:

# ✓ GOOD: Explains WHY
TIMEOUT = 30        # generous — API flaky on Mondays
MAX_RETRIES = 3     # based on observed failure patterns
BATCH_SIZE = 100    # memory-safe for 8GB instances

# ✓ GOOD: Documents intent
# TODO: Add circuit breaker after next outage
# NOTE: Duplicates logic in api_client.py — consider extracting

The rule: Every comment token should carry meaning. If it's just decoration, delete it.

Argparse vs Click vs Typer

Argparse is preferred for sniffable Python because it separates structure from implementation:

The Argparse Advantage

With Click/Typer, decorators attach to functions — so CLI structure is interleaved with implementation:

# Click: structure mixed with implementation
@cli.command()
def examine(target: str):  # ← implementation starts here
    """Look at something."""
    do_stuff()  # ← can't sniff without reading this

With argparse, main() contains the entire CLI tree with no implementation:

# Argparse: clean separation
def main():
    """CLI structure only — no implementation."""
    parser = argparse.ArgumentParser(description="Tool description")
    subparsers = parser.add_subparsers(dest="command")

    # examine command
    examine_parser = subparsers.add_parser("examine", help="Look at something")
    examine_parser.add_argument("target", help="What to examine")
    examine_parser.add_argument("--verbose", "-v", action="store_true")

    # move command
    move_parser = subparsers.add_parser("move", help="Move in a direction")
    move_parser.add_argument("direction", choices=VALID_DIRECTIONS)

    args = parser.parse_args()
    _dispatch(args)  # ← implementation is ELSEWHERE

# IMPLEMENTATION (completely separate, below the fold)
def _dispatch(args):
    ...

The LLM reads main() and sees the entire command tree — arguments, types, choices, help text — without any implementation noise. Pure sniffability. No need to hold your nose.

Integration with Sister Scripts

Sniffable Python is how sister scripts should be structured:

skill/
├── SKILL.md           # Tells LLM how to generate scripts
├── CARD.yml           # Interface definition
└── scripts/
    └── helper.py      # Sniffable Python

When the skill/SKILL.md instructs the LLM to generate a sister script, it should follow sniffable-python conventions:

1. Docstring with usage at top, no decorations
2. Imports grouped and labeled
3. Constants with explanatory comments
4. CLI structure before implementation
5. Implementation below the fold

Directory-Agnostic Invocation

Critical pattern: All scripts — including sniffable Python — must work when invoked from ANY directory. Scripts find their skill context using __file__, not os.getcwd().

For bash scripts, see sister-script/ for the equivalent pattern.

The Pattern

from pathlib import Path

# Find THIS script's location (not the caller's cwd)
SCRIPT_DIR = Path(__file__).resolve().parent
SKILL_DIR = SCRIPT_DIR.parent  # If script is in skill/scripts/

# Now find sibling files relative to script location
CONFIG_FILE = SKILL_DIR / "config.yml"
PATTERNS_DIR = SKILL_DIR / "patterns"
TEMPLATES_DIR = SKILL_DIR / "templates"

Why __file__?

• Path(__file__) — path to THIS script file
• .resolve() — resolve symlinks to get real location
• .parent — containing directory

Why NOT os.getcwd()?

• os.getcwd() returns the CALLER's directory
• Script invoked from /home/user/ won't find skills/foo/patterns/

The Complete Template

#!/usr/bin/env python3
"""sister-tool: Brief description.

Usage:
    python3 skills/my-skill/scripts/sister-tool.py command [options]

Can be invoked from any directory — finds its own context.
"""

from pathlib import Path
import argparse
import yaml

# DIRECTORY CONTEXT
# Find script location regardless of caller's cwd
SCRIPT_DIR = Path(__file__).resolve().parent
SKILL_DIR = SCRIPT_DIR.parent  # scripts/ is one level down from skill/

# Skill-relative paths
CARD_FILE = SKILL_DIR / "CARD.yml"
PATTERNS_DIR = SKILL_DIR / "patterns"
TEMPLATES_DIR = SKILL_DIR / "templates"
CONFIG_DIR = SKILL_DIR / ".moollm" / "skills" / SKILL_DIR.name

# CONFIGURATION
def load_card():
    """Load skill's CARD.yml for metadata."""
    if CARD_FILE.exists():
        return yaml.safe_load(CARD_FILE.read_text())
    return {}

def main():
    """CLI structure — sniffable, directory-agnostic."""
    card = load_card()
    parser = argparse.ArgumentParser(
        description=card.get("description", __doc__.split('\n')[0])
    )
    # ... CLI structure ...
    args = parser.parse_args()
    _dispatch(args)

# IMPLEMENTATION
def _dispatch(args):
    ...

if __name__ == "__main__":
    main()

Why This Matters

Path(__file__).parent always works. The script finds its own context.

The Play-Learn-Lift Connection

Sniffable Python is the crystallization point of LIFT — where proven procedures become reusable automation.

# PLAY → LEARN → LIFT progression
play_learn_lift:
  play:
    emoji: "🎮"
    action: "Jump in! Try things"
    artifact: "session-log.md"
    quality: "Messy exploration, may fail"
  learn:
    emoji: "📚"
    action: "Patterns emerge, document procedures"
    artifact: "PROCEDURE.md"
    quality: "Structured notes, works when followed"
  lift:
    emoji: "🚀"
    action: "Automate it! Share the tool"
    artifact: "sister-script.py (SNIFFABLE PYTHON)"
    quality: "Clean interface, others can use"

The LIFT stage produces sniffable Python because:

• Others (human and LLM) need to understand the tool quickly
• The CLI must be discoverable without reading implementation
• The script must be maintainable without spelunking

See: play-learn-lift/ — The full methodology

The Skill-to-Script Flow

The skill/ skill describes how skills act as factories that produce instances. When a skill generates a CLI tool, that tool should be sniffable Python.

# Skill → Script → LLM → Output pipeline
skill_to_script_flow:
  - stage: "SKILL.md"
    role: "The teacher — knows how to build things"
    action: "instructs LLM to generate"
  - stage: "Sister Script"
    role: "SNIFFABLE PYTHON"
    properties:
      - "Docstring = API"
      - "main() = CLI tree"
      - "Comments = Jazz"
    action: "LLM sniffs to understand"
  - stage: "LLM Comprehension"
    role: "Reads first 50 lines, understands CLI, can invoke tool"
    action: "executes"
  - stage: "Structured Output"
    role: "YAML / JSON — LLM can parse, feeds back"

Example from adventure skill:

The adventure skill's vision includes a Python CLI that handles deterministic operations:

$ adventure lint examples/adventure-4/    # Validate world
$ adventure move alice north              # Update coordinates
$ adventure scan --pending                # Find work to do

Each of these is a sniffable subcommand. The LLM sniffs adventure --help (or just reads main()) and knows the full API.

The Adventure Linter Feedback Loop

The adventure/ skill demonstrates sniffable Python in a feedback loop:

# Adventure linter feedback loop
linter_feedback_loop:
  - step: "LLM generates content"
    does: "Create rooms, objects, characters in YAML Jazz format"
  - step: "Sniffable linter runs"
    does: "adventure-lint.py validates world consistency"
  - step: "LINTER.yml output"
    does: "Structured findings: errors, warnings, suggestions"
  - step: "LLM reads lint results"
    does: "Understands what needs fixing, generates corrections"
  - step: "iterate"
    does: "Loop back to generation with fixes"

From examples/adventure-4/LINTER.yml:

summary:
  rooms_found: 36
  objects_found: 37
  characters_found: 6
  errors: 8
  warnings: 36
  compile_requests: 9

events:
- type: FOUND_ROOM
  severity: INFO
  path: pub/ROOM.yml
  message: 'Found room: The Gezelligheid Grotto'
  data:
    exits: [NORTH, UP, BACK]
    objects: 15
    sub_rooms: 1

Why this works:

• The linter is sniffable — LLM understands its CLI
• The output is YAML — LLM parses it naturally
• Events are structured — LLM knows what to fix
• The loop is tight — generate → lint → fix → repeat

This is Python for precision, LLM for poetry.

Live Examples in adventure-4

TomTomagotchi — YAML Jazz in Action

kitchen/tomtomagotchi.yml demonstrates sniffable structure in YAML:

abilities:
  COMPASS:
    description: "Point toward a target"
    usage: "COMPASS [target]"
    targets:
      - start     # Always knows home
      - home      # The surface
      - treasure  # The goal

The comment "Always knows home" is data. The LLM reads it and understands the semantics. This is YAML Jazz applied to game objects.

The Pub — Framing Through Comments

pub/ROOM.yml uses comments to frame the room's purpose:

framing:
  mode: [performance, celebration, socialization, tribute, third_place]

  description: |
    This is a PLACE OF PERFORMANCE AND CELEBRATION.
    This is a THIRD PLACE — neither home nor work, but community.
    This is where TRIBUTES are performed with love.

Comments explain WHY the pub exists. The LLM uses this framing to generate appropriate content — debates are sport, personas are welcome, tributes are loving.

Skills as Rooms

From skill/SKILL.md, skills manifest as rooms you can enter:

> enter the adventure skill
You are in the Adventure Workshop.
Exits: pub, maze, character-gallery
Objects: room-templates, npc-catalog, puzzle-designs

The skill directory IS the room. The sister scripts are objects you can pick up and use. Sniffable Python makes those scripts discoverable.

The Coherence Engine Flow

# Coherence Engine: user request → LLM discovery → execution
coherence_flow:
  trigger: "Run the room builder for the kitchen"
  steps:
    - action: "LLM reads skill/SKILL.md"
      discovers: "scripts/room-builder"
    - action: "LLM sniffs room-builder.py (first 50 lines)"
      sees:
        docstring: "Creates rooms..."
        commands: [create, modify, list]
        options: [--template, --force]
    - action: "LLM understands API, generates command"
      output: "room-builder create --template kitchen"

The LLM discovers the tool's capabilities by reading its structure, not by loading documentation. One quick sniff and it knows what's cooking.

Checklist

When writing sniffable Python:

• Shebang on line 1
• Module docstring with purpose, usage, examples
• Imports grouped at top (no decorative markers)
• Constants with explanatory comments
• CLI structure in main() using argparse (preferred) or Click/Typer
• Each command has a docstring
• Types specified in function signatures
• Implementation below line 50
• Internal functions prefixed with _
• if __name__ == "__main__": at bottom

Anti-Patterns (Code Smells You Can't Unsmell)

❌ Implementation before CLI

def _helper():
    ...
def _another_helper():
    ...
# 200 lines later...
@click.command()
def main():
    ...

❌ No docstrings

@click.command()
def process(x, y, z):
    # What does this do? What are x, y, z?

❌ Magic constants

if retries > 3:  # Why 3? What happens at 4?

❌ Scattered imports

import os
def foo():
    import json  # Hidden dependency

Commands

SNIFF CODE — Works on Any Language

The SNIFF CODE command isn't Python-specific. It reviews any source file:

> SNIFF CODE src/api/handler.ts

Sniffability Report for handler.ts:
✓ Purpose clear from first 10 lines
✓ Exports/API visible before implementation
✗ Magic number on line 47 — needs comment
✗ Helper function before main export — reorder
✓ Semantic comments explain WHY not WHAT

Overall: Mostly fresh, minor reordering needed

What it checks:

Good for humans. Good for LLMs. Same conventions.

Protocol Symbol

SNIFFABLE-PYTHON

Invoke when: Generating or reading Python scripts that need to be LLM-comprehensible.

Dovetails With — The Intertwingularity

Everything in MOOLLM connects. Sniffable Python sits at the intersection of multiple skills:

# How sniffable-python connects to other skills
skill_connections:
  parent: "constructionism (build to learn)"
  center: "sniffable-python"
  peers:
    - skill: "play-learn-lift"
      relation: "LIFT stage produces sniffable Python"
    - skill: "yaml-jazz"
      relation: "Comments carry semantic meaning"
  children:
    - skill: "sister-script"
      relation: "Sister scripts ARE sniffable Python"
    - skill: "skill"
      relation: "Skills generate sniffable scripts"
    - skill: "adventure"
      relation: "Linter exemplifies the feedback loop"

Live Examples

The Circle Completes

PLAY: Try things, log them
  ↓
LEARN: Document patterns in PROCEDURE.md
  ↓
LIFT: Generate sister-script.py in SNIFFABLE PYTHON format
  ↓
SKILL: Add to skill, expose to LLM
  ↓
LLM: Sniffs script, understands API, uses tool
  ↓
OUTPUT: YAML results (linter, builder, etc.)
  ↓
PLAY: Try new things with the tool...
  ↓
[repeat]

This is the MOOLLM development loop. Sniffable Python is the code interface that makes it work.

Beyond Python: Universal Sniffability

While this skill focuses on Python, the principles are language-agnostic:

The universal rule: Whatever a reader (human or LLM) needs to *use* the code should be visible before whatever they need to *modify* it.

// ✓ Sniffable TypeScript
/**
 * User authentication service.
 * @module auth
 */

export interface AuthConfig { ... }
export function authenticate(config: AuthConfig): Promise<User> { ... }
export function logout(): void { ... }

// Implementation details below
function validateToken(token: string): boolean { ... }

Same smell test, any language: Can you understand what this code offers by reading just the head?

Lickable Pixels, Sniffable Code

When Steve Jobs unveiled Mac OS X's Aqua interface in 2000, he said:

*"We made the buttons on the screen look so good you'll want to lick them."*

Jobs understood that interfaces aren't just functional — they should be sensually inviting. The translucent droplets, the pulsing default buttons, the candy-colored icons — these weren't decoration. They were signals of quality that drew users in.

Programming languages are user interfaces. The users are humans and LLMs. And just as lickable pixels invite visual exploration, sniffable code invites structural exploration.

Jobs didn't make Aqua pretty at the expense of function — the visual design *reinforced* usability. Sniffable code works the same way: the structural clarity that helps LLMs also helps humans. We're not optimizing for machines at human expense. We're optimizing for comprehension, and both benefit.

*Perl looks like line noise. Sniffable Python looks like an invitation.*

The Thesis

The "ideal LLM syntax" question has the wrong framing.

You don't need new syntax. You need structured familiar syntax.

• Languages LLMs know (Python, TypeScript, Go, Rust...)
• Structured for sniffability (important stuff at top)
• Comments as data (YAML Jazz applied to any language)
• Single source of truth (the code IS the spec)

Dense syntax compresses the wrong thing. Sniffable code structures the right thing.

*Jobs made you want to lick the screen. We want you to sniff the source.*

*"The LLM doesn't need fewer tokens of unfamiliar syntax. It needs familiar syntax, structured for fast comprehension."*

*Good code doesn't just avoid bad smells — it has an aroma that draws you in from the header.*