skill
stan-skills
Agent Skills for Bayesian modeling in Stan, informed by Gelman/Vehtari/McElreath's Bayesian Workflow and Regression and Other Stories
Set up
Skills are installed by placing their folder where your agent loads skills.
The download comes from github.com/jacobpstein/stan-skills itself, not from our server, so what you get is exactly what its author published.
Install instructions come from the README
README
synced Oct 23, 2026Stan Skills
A plugin for Claude Code and other AI coding agents providing Agent Skills for Bayesian modeling in Stan, informed by the workflow and applied-regression guidance in
- Gelman, Vehtari, McElreath et al., Bayesian Workflow (CRC Press, 2026), and
- Gelman, Hill, Vehtari, Regression and Other Stories (Cambridge, 2020).
The skills cover the Python interface (CmdStanPy + ArviZ) and the R interfaces (cmdstanr, rstanarm, brms, posterior, bayesplot, loo) on equal footing. The structure mirrors pymc-labs/python-analytics-skills.
Skills
| Skill | Description |
|---|---|
| bayesian-workflow | The full Gelman/Vehtari/McElreath workflow: pick an initial model, prior predictive check, fake-data simulation, fit and diagnose, posterior predictive check, cross-validation, expand one step at a time, compare, and report the series of models. |
| stan-modeling | Writing Stan programs: blocks, current syntax, types and constraints, vectorization, non-centered parameterization, GLMs, hierarchical models, GPs, time series, mixtures, censoring, ordinal, measurement error; fitting with CmdStanPy/cmdstanr and moving draws into ArviZ/posterior. |
| stan-priors | Prior specification: weakly informative defaults and scaling rules, priors for regression/multilevel/GLMM models, prior predictive checks, tail behavior and prior-likelihood conflict, power-scaling sensitivity (priorsense), regularizing and sparsity priors. |
| stan-computation | Fitting and diagnosing HMC/NUTS: warmup and adaptation, chains and iterations, R-hat/ESS/MCSE thresholds, divergences and treedepth, the fail-fast loop, named failure modes with fixes, reparameterization, multimodality, approximate algorithms (Laplace, Pathfinder, ADVI), and parallelization. |
| stan-model-evaluation | Model checking and comparison: posterior predictive checks and test statistics, PSIS-LOO and Pareto k, moment matching, k-fold, integrated LOO, LOO-PIT calibration, loo_compare, stacking, Bayesian R², projection predictive variable selection. |
| applied-regression | Regression and Other Stories in Stan/rstanarm: coefficients as comparisons, divide-by-4, centering/standardizing, log transformations, interactions, GLMs (logistic, Poisson, negative binomial, ordered), fake-data simulation, design analysis, poststratification, and causal regression pitfalls. |
| r-stan-interfaces | The R stack: cmdstanr, rstanarm (stan_glm, stan_glmer, default priors), brms (brm, set_prior, make_stancode), posterior, bayesplot, loo, priorsense, projpred. |
| stan-testing | Statistical modeling as software development: parameter-recovery tests, simulation-based calibration (SBC), pytest/testthat harnesses for Stan programs, fixed_param and generate_quantities for fast tests, reproducibility and version control. |
Installation
Via npx
npx skills add jacobpstein/stan-skills
As a Claude Code plugin
/plugin marketplace add jacobpstein/stan-skills
/plugin install stan@stan-skills
Installs all skills plus the keyword-suggestion hook.
Manual installation
git clone https://github.com/jacobpstein/stan-skills.git
cd stan-skills
./install.sh claude # Claude Code
./install.sh all # All platforms
./install.sh claude -- stan-modeling # A single skill
Utility commands
./install.sh --list # List skills with descriptions
./install.sh --validate # Validate skill structure
./scripts/validate-skills.sh
Platform support
| Platform | Install location | Auto-discovered |
|---|---|---|
| Claude Code | ~/.claude/skills/ | Yes |
| OpenCode | ~/.config/opencode/skills/ | Yes |
| Gemini CLI | ~/.gemini/skills/ | Yes |
| Cursor | ~/.cursor/skills/ | Yes |
| VS Code Copilot | ~/.copilot/skills/ | Yes |
Project structure
stan-skills/
├── .claude-plugin/
│ ├── marketplace.json # Plugin registry metadata
│ ├── plugin.json # Plugin configuration
│ └── CLAUDE.md # Plugin development guidance
├── hooks/
│ ├── hooks.json # UserPromptSubmit hook configuration
│ └── suggest-skill.sh # Keyword-based skill suggestion
├── scripts/
│ └── validate-skills.sh # Repository validation
├── skills/
│ ├── bayesian-workflow/
│ ├── stan-modeling/
│ ├── stan-priors/
│ ├── stan-computation/
│ ├── stan-model-evaluation/
│ ├── applied-regression/
│ ├── r-stan-interfaces/
│ └── stan-testing/
├── install.sh
├── package.json
└── skills.json
Each skill is a SKILL.md entry point plus references/*.md for detail the agent loads on demand.
Hooks
A UserPromptSubmit hook suggests relevant skills when it detects keywords such as stan, cmdstanpy, cmdstanr, rstanarm, brms, divergences, R-hat, loo, Pareto k, prior predictive, fake data, poststratification, SBC. Test it directly:
echo '{"prompt": "my Stan model has divergences"}' | bash hooks/suggest-skill.sh
Sources
The workflow content follows Bayesian Workflow (Gelman, Vehtari, McElreath, with Simpson, Margossian, Yao, Kennedy, Gabry, Bürkner, Modrák, Leos Barajas; CRC Press 2026; code at https://avehtari.github.io/Bayesian-Workflow/) and Regression and Other Stories (Gelman, Hill, Vehtari; Cambridge 2020; code at https://avehtari.github.io/ROS-Examples/). The skills paraphrase and operationalize the guidance; they do not reproduce book text.
Contributing
See CONTRIBUTING.md.
License
MIT. See LICENSE.