Implementing Code

Core philosophy: Write it right the first time — clean, concise, conceptually dense code that doesn't need editing afterward. Zero linting violations (ruff or project-specified tools). Check CLAUDE.md for project-specific standards.

Conceptual Density

Fit as many related operations into one line as possible, where each line has an understandable big-picture purpose. Each line = one complete concept. 88 char max (ruff default).

# Inline calculations in dict construction — not scattered assignments
results = {
    "chi2_E": hartlap_factor * (En @ np.linalg.solve(cov_E, En)),
    "pte_B": 1 - stats.chi2.cdf(chi2_B, nmodes),
    "cov": np.cov(np.array(samples).T),
}

# Eliminate repetition with comprehensions
chi_E, chi_B, chi_EB = [
    np.sum(signal**2 / noise**2)
    for signal, noise in zip((ee, bb, eb), (noise_ee, noise_bb, noise_eb))
]

# Consolidate repeated plotting blocks into loops
for ax, (cl, err, fmt, label) in zip(axes, plot_data):
    ax.errorbar(ell_bins, cl, yerr=err, fmt=fmt)
    ax.set(xlabel=r"$\ell$", ylabel=label, title=f"{label} Power Spectrum")

Concise conditionals

# One-line conditionals over verbose if/else blocks
n_eff = n_samples if cov_path_int is not None else npatch
version_results = results_list[idx] or self.calculate_pure_eb()
output_path = user_path or generate_default_path()

# Short-circuit execution
(var_method == "semianalytic") and self.calculate_semianalytic_cov()

Natural line breaking

When code exceeds 88 chars, break at logical boundaries:

Comprehensions: after [ and before for
Function calls: at commas between argument groups
Dicts: one key-value pair per line
Chained operations: at method calls or operators

Anti-Patterns

Unnecessary intermediates: temp = solve(A, b); result = factor * temp -- combine into one conceptual line
Scattered dict assignments: results["a"] = a; results["b"] = b -- use dict construction with inline calculations
Verbose conditionals: 4-line if/else for simple assignment -- use ternary or or
Defensive .get() with defaults: config.get("key", default) hides missing config -- use config["key"] to fail fast
Unnecessary try/except: let errors propagate; missing files and bad inputs are real problems
Session-specific comments: "per your request", "switched to new method" -- comments should be timeless

Comment Philosophy

Comments explain why and context, never what. Good: # Hartlap correction for finite sample bias. Bad: # Create numpy array. Remove artifacts like "change to scipy implementation" from library code.

Error Handling

Trust scientific libraries to validate their domains. Skip defensive programming. Let errors propagate -- failed fast is better than hidden bugs. Use defensive patterns only for known edge cases with clear scientific justification.

Validation

Spot-check that code runs without crashing
Test with toy data where feasible
Extended validation (literature comparison, edge cases) only when requested
Focus on "does it work" over exhaustive test suites

Working Agreements

Before implementing numerical algorithms, think through stability and convergence
Deliver working code, basic validation, and integration notes (if connecting to existing code)