PyMC Modeling Skill for Claude Code

A Claude Code skill that provides comprehensive guidance for Bayesian statistical modeling with PyMC v5+. When loaded, Claude gains expert knowledge of PyMC workflows, best practices, and common patterns.

When you work on PyMC modeling tasks, this skill is automatically loaded to provide:

• Modern PyMC v5+ API patterns
• MCMC inference best practices (nutpie, NumPyro/JAX backends)
• ArviZ diagnostic workflows
• Prior specification guidance
• Common model templates (hierarchical, GLMs, GPs, time series, BART, mixtures)

Installation

Copy or symlink this directory to your Claude Code skills folder:

# Create skills directory if it doesn't exist
mkdir -p ~/.claude/skills

# Clone or copy this repository
git clone https://github.com/fonnesbeck/pymc-modeling.git ~/.claude/skills/pymc-modeling

The skill will automatically activate when Claude detects tasks involving Bayesian inference, PyMC, ArviZ, or related topics.

Structure

SKILL.md                      # Main skill: model specification, inference, diagnostics
CLAUDE.md                     # Instructions for Claude when editing this skill
references/
  arviz.md                    # Comprehensive ArviZ guide (all plots, interpretation)
  priors.md                   # Prior selection (weakly informative defaults)
  inference.md                # MCMC backends (nutpie, NumPyro/JAX, approximate)
  diagnostics.md              # Quick diagnostics reference, troubleshooting
  gotchas.md                  # Common pitfalls and performance issues
  gp.md                       # Gaussian processes (HSGP, kernels, priors)
  timeseries.md               # Time series (AR, structural, seasonality)
  bart.md                     # BART (pymc-bart) usage and interpretation
  mixtures.md                 # Mixture models and label switching
  specialized_likelihoods.md  # Zero-inflated, censored, ordinal, robust
  custom_models.md            # Custom distributions (DensityDist, Potential)

Topics Covered

Core Workflow

• Model specification with coords/dims for interpretable InferenceData
• Centered vs non-centered parameterization
• Inference with nutpie (default), PyMC NUTS, and NumPyro/JAX
• Systematic 5-phase ArviZ diagnostic workflow
• Prior and posterior predictive checks
• Model comparison (LOO-CV, WAIC)
• Saving/loading results

Model Types

• Hierarchical/multilevel models: Partial pooling, non-centered parameterization
• GLMs: Logistic, Poisson, negative binomial regression
• Gaussian processes: HSGP for scalability, periodic GPs, kernel selection
• Time series: AR, random walk, structural time series, seasonality
• BART: Bayesian additive regression trees via pymc-bart
• Mixture models: Gaussian mixtures, label switching solutions
• Specialized likelihoods: Zero-inflated, hurdle, censored, ordinal, robust

Advanced Topics

• Custom distributions with pm.DensityDist and pm.CustomDist
• Soft constraints with pm.Potential
• Simulation-based inference with pm.Simulator
• Causal inference with pm.do and pm.observe
• pymc-extras for marginalized mixtures and R2D2 priors

Key Recommendations

When this skill is loaded, Claude will follow these principles:

Topic	Recommendation
Sampler	nutpie by default (2-5x faster), PyMC NUTS as fallback
Parameterization	Non-centered for hierarchical models with weak data
GPs	HSGP for n > 500 points; full GP only for small datasets
Diagnostics	Check r_hat < 1.01, ESS > 400, zero divergences before interpretation
Notebooks	marimo preferred over Jupyter

Requirements

The skill assumes these packages are available:

• pymc >= 5.0
• arviz
• nutpie (recommended)
• pymc-bart (for BART models)
• pymc-extras (for specialized distributions)

Optional for GPU/JAX acceleration:

• numpyro
• jax

License

This skill is provided as-is for use with Claude Code.