fastaistyle

pip install fastaistyle

A style checker that enforces the fast.ai coding style—a compact, readable approach to Python that keeps more code visible on screen and reduces cognitive load.

Why This Style?

Most style guides optimize for the wrong thing. They add vertical space, mandate verbose names, and scatter related code across many lines. The result? You see less code at once, which means more scrolling, more context-switching, and more mental effort to understand what's happening.

The fast.ai style takes a different approach, rooted in decades of experience with APL, J, K, and scientific programming. The core insight comes from Kenneth Iverson: "brevity facilitates reasoning."

Your Brain Can Only Hold So Much

When you're reading code, your working memory is limited. If a function spans 50 lines, you can't see the whole thing at once. You scroll down, forget what was at the top, scroll back up. Each scroll is a context switch. Each context switch costs mental energy.

But if that same function fits in 15 lines? You see the whole picture. Your eyes can jump between related parts instantly. Patterns become obvious. Bugs stand out.

This isn't about cramming code together—it's about removing unnecessary vertical space so your brain can do what it's good at: recognizing patterns across visible information.

One Line, One Idea

The goal is density without confusion. Each line should express one complete thought:

# Good: you see the whole pattern at once
if not data: return None
for item in items: process(item)
def _is_ready(self): return self._ready.is_set()

# Bad: same logic, but now it's 6 lines instead of 3
if not data:
    return None
for item in items:
    process(item)
def _is_ready(self):
    return self._ready.is_set()

When the body is simple, keep it on the same line. Save vertical space for code that actually needs it.

Names Should Be Short (When Used Often)

This follows "Huffman coding" for variable names—frequently used things get short names:

# Good: conventional, recognizable
img, i, msg, ctx

# Bad: verbose for no benefit
image_data, loop_index, message_object, context_instance

Domain experts recognize nll (negative log likelihood) instantly. Spelling it out doesn't help them, and the extra characters push code off the right edge of the screen.

Installation

pip install fastaistyle

Or install from source:

git clone https://github.com/AnswerDotAI/fastaistyle
cd fastaistyle
pip install -e .

Usage

Check the current directory:

chkstyle

Check a specific path:

chkstyle path/to/code/

Skip folders matching a regex (must match the whole folder name):

chkstyle --skip-folder-re 'test.*|migrations|vendor'

The checker prints violations with file paths, line numbers, and the offending code.

Configuration

Configure chkstyle in your pyproject.toml:

[tool.chkstyle]
skip-folder-re = "test.*|migrations|vendor"

Command-line arguments override config file settings.

What It Checks

dict literal with 3+ identifier keys

Use dict() for keyword-like keys—it's easier to scan and produces cleaner diffs.

# Bad
payload = {"host": host, "port": port, "timeout": timeout}

# Good
payload = dict(host=host, port=port, timeout=timeout)

single-statement body not one-liner

If the body is one simple statement, keep it on the header line.

# Bad
if ready:
    return True

# Good
if ready: return True

single-line docstring uses triple quotes

Triple quotes are for multi-line strings. Single-line docstrings should use regular quotes.

# Bad
def foo():
    """Return the value."""
    return x

# Good
def foo():
    "Return the value."
    return x

multi-line from-import

If it fits on one line, keep it on one line.

# Bad
from os import (
    path,
    environ,
)

# Good
from os import path, environ

line >150 chars

Wrap at a natural boundary: argument lists, binary operators, or strings. But 150 is generous—aim for ~120 when practical.

semicolon statement separator

Don't use ; to combine statements. Use separate lines.

inefficient multiline expression

If the content would fit in fewer lines, condense it.

# Bad
result = call(
    a,
    b,
    c,
)

# Good
result = call(a, b, c)

lhs assignment annotation

Avoid x: int = 1 in normal code. Put type hints on function parameters and return values instead. The exception is dataclass fields, where annotations are required.

# Bad
x: int = 1
name: str = "hello"

# Good (in a function signature)
def process(x: int, name: str) -> Result: ...

nested generics depth >= 2

Keep type annotations simple. Deep nesting makes them hard to read.

# Bad
items: list[dict[str, list[int]]]

# Good
Payload = dict[str, list[int]]
items: list[Payload]

Opting Out

Sometimes you have a good reason to format code a specific way. The checker supports pragmas:

# Ignore a single line
x: int = 1  # chkstyle: ignore

# Ignore the next line
# chkstyle: ignore
y: int = 2

# Disable for a block
# chkstyle: off
carefully_formatted = {
    "alignment": "matters",
    "here":      "for readability",
}
# chkstyle: on

# Skip an entire file (must be in first 5 lines)
# chkstyle: skip

Running Tests

pytest

Development

Create a PR with a label (for release notes):

./tools/pr.sh enhancement "Add new feature"
./tools/pr.sh bug "Fix something"

Release after PRs are merged:

./tools/release.sh        # patch bump (default)
./tools/release.sh minor  # minor bump
./tools/release.sh major  # major bump

This tags the current version, pushes to trigger PyPI publish, then bumps for the next dev cycle.

Philosophy

This tool exists to catch mechanical issues, not to enforce taste. The violations it reports are almost always things you'd want to fix—extra vertical space that doesn't help, type annotations that clutter rather than clarify, formatting that makes diffs noisier than necessary.

The goal is code that's pleasant to read and easy to maintain. Dense, but not cramped. Clear, but not verbose. When in doubt, look at the surrounding code and match its style.

For the full style guide, see:

• fast.ai Style Guide
• style.md in this repo

License

Apache 2.0