Clean Code Agent Skill

Clean Code

Write code that humans can read, understand, and change with confidence.

Core Principles

1. Mean What You Say — Say What You Mean

Code is communication. Every name, structure, and abstraction should reveal intent. If you need a comment to explain *what* the code does, the code is failing at communication.

2. Small Things, Done Well

Small functions, small classes, small files. Each unit of code should have one clear responsibility and do it well. Composability beats complexity.

3. The Boy Scout Rule

Leave the code cleaner than you found it. Every commit should improve the codebase incrementally — even if it's just renaming one variable or extracting one function.

4. Testability == Design Quality

If code is hard to test, it has a design problem. Testable code is modular, decoupled, and honest about its dependencies.

Clean Code Scoring Rubric

Use this rubric to evaluate code quality on a scale of 1-5 for each dimension:

| Dimension | 1 (Poor) | 3 (Adequate) | 5 (Excellent) |

|-----------|----------|---------------|----------------|

| Naming | Single-letter vars, ambiguous abbreviations | Descriptive but occasionally redundant | Reveals intent, consistent, searchable |

| Function Size | Monolithic 500+ line functions | 50-100 line functions with mixed concerns | <20 lines, one clear level of abstraction |

| Comments | Outdated or redundant comments | Comments explain *what* not *why* | Minimal comments, code is self-documenting |

| Error Handling | Silent catches, magic error codes | Basic try/catch, some error types | Rich error types, graceful degradation |

| Testing | No tests or brittle tests | Tests exist but tightly coupled to implementation | Tests specify behavior, not implementation |

| Duplication | Copy-paste everywhere | Some reuse, some DRY violations | DRY with well-abstracted patterns |

Target: 4+ in every dimension for production-grade code.

Actionable Guidance

Naming

Rules:

• Boolean variables: Use positive names (isActive, hasPermission, shouldRetry). Avoid negated names like isNotDisabled.
• Functions/methods: Verbs or verb phrases (calculateTotal(), validateInput(), fetchUser()).
• Classes/types: Nouns or noun phrases (UserAccount, PaymentProcessor, HttpClient).
• Constants: UPPER_SNAKE_CASE (MAX_RETRY_COUNT, DEFAULT_TIMEOUT_MS).

# Bad
def proc(d):
    r = []
    for i in d:
        if i.get('a') == True:
            r.append(i.get('n'))
    return r

# Good
def extract_active_user_names(users):
    active_users = [user for user in users if user['is_active']]
    return [user['name'] for user in active_users]

Searchable names: Avoid single-letter variables except in trivial loops. Use names that can be found with grep.

Functions

Rules:

• One level of abstraction per function: A function should mix high-level logic (e.g., "fetch data") with mid-level logic (e.g., "parse CSV line") or low-level (e.g., "trim whitespace") — never all three.
• 3-4 parameters max: More than 4 suggests the function does too much. Bundle related params into objects.
• No side effects: Prefer pure functions. If a function must mutate state, make it obvious (name it setX(), updateY()).
• DRY but not at cost of clarity: Extract duplication into shared helpers, but don't create overly abstracted indirection for code that appears only twice.

// Bad: Mixed abstraction levels
function processOrder(order) {
  const tax = order.total * 0.08;                       // Low-level calc
  order.totalWithTax = order.total + tax;                // Mutation
  fs.writeFileSync(`orders/${order.id}.json`, JSON.stringify(order)); // Side effect
  sendEmailNotification(order.userEmail, 'Order processed'); // Side effect
  return order.totalWithTax;
}

// Good: Clear single responsibility
function calculateTotalWithTax(total, taxRate) {
  return total + (total * taxRate);
}

Comments

When to comment:

• Tricky business logic: Why a specific algorithm was chosen
• Non-obvious tradeoffs: Why you chose A over B
• Legal/copyright: Required attribution
• Warnings: // FIXME: This endpoint is rate-limited to 100 req/min

When NOT to comment:

• Stating the obvious (// Increment counter)
• Commenting-out dead code — delete it. Git remembers.
• Writing a novel — if you need paragraphs, your code needs refactoring.

Error Handling

Patterns:

# Prefer specific exception types
def get_user(user_id):
    try:
        return database.fetch_user(user_id)
    except DatabaseConnectionError:
        logger.error(f"Database unavailable when fetching user {user_id}")
        raise ServiceUnavailableError("User service temporarily unavailable")
    except UserNotFoundError:
        logger.info(f"User {user_id} not found")
        return None  # Expected case, not exceptional

Guidelines:

• Fail fast: Validate inputs at boundaries. Don't let bad data propagate.
• Return typed errors: Use Result[T, E] types (Rust, Swift) or Either (functional languages) instead of exceptions for expected failures.
• Never swallow exceptions: Empty catch blocks are a code smell. At minimum, log and re-raise.
• Use error codes sparingly: HTTP status codes make sense at API boundaries. Inside your application, use typed errors.

Testing

The Testing Trophy (not pyramid):

     E2E Tests  (few)
    Integration (some)
   Unit Tests   (many)
  Static Analysis (all code)

Guidelines:

• Test behavior, not implementation. Your tests should pass after a refactor if the behavior didn't change.
• One assertion concept per test. Use multiple it() blocks rather than multiple asserts in one.
• Use realistic test data. "foo" and 123 don't catch edge cases.
• For AI-generated code: always verify with tests. The AI writes the code, you write the tests.

# Bad: Tests implementation details
def test_get_user():
    mock_db = MagicMock()
    service = UserService(mock_db)
    result = service._fetch_and_transform_user(42)  # Testing private method
    assert mock_db.execute.called_once_with("SELECT * FROM users WHERE id=42")

# Good: Tests behavior
def test_get_user_returns_user_when_found():
    user_repo = InMemoryUserRepository([User(id=42, name="Alice")])
    service = UserService(user_repo)
    result = service.get_user(42)
    assert result.name == "Alice"

def test_get_user_returns_none_when_not_found():
    user_repo = InMemoryUserRepository([])
    service = UserService(user_repo)
    result = service.get_user(99)
    assert result is None

Code Smells to Hunt

| Smell | Symptom | Fix |

|-------|---------|-----|

| Long Method | >20 lines doing multiple things | Extract methods, compose |

| Switch/Types | Switch on type enum, then dispatch | Polymorphism or strategy pattern |

| Feature Envy | Method uses more of another class's data than its own | Move method to the right class |

| Shotgun Surgery | One change requires edits in many files | Consolidate related logic |

| Data Clumps | Same 3-4 fields appear together repeatedly | Extract into a value object |

| Primitive Obsession | Using strings/ints where types belong | Create domain types |

| Inappropriate Intimacy | Class knows too much about another's internals | Reduce coupling, use interfaces |

Common Mistakes

• Over-optimizing for performance before clarity: 99% of code doesn't need micro-optimization. Write clear code first, profile, then optimize the hot paths.
• Over-engineering: YAGNI (You Ain't Gonna Need It). Don't add abstractions for hypothetical future needs.
• Perfect as enemy of good: Clean code is a journey, not a destination. Incremental improvement beats paralysis.
• Ignoring the team's conventions: Consistency within a codebase matters more than personal preference for a particular style.
• Applying rules blindly: All rules have exceptions. goto in C error handling is fine. Single-letter variables in math-heavy code are fine. Context matters.