Bayesian Statistician Plugin

A Claude Code plugin for end-to-end Bayesian statistical modeling with Stan and ArviZ. It packages an orchestrator skill, eleven specialized subagents, and a library of modeling skills that together run a full Bayesian workflow: EDA → model design → fitting → validation → reporting.

Prerequisites

• uv — used for all Python execution.
• A C++ toolchain — CmdStanPy compiles Stan models against CmdStan. The /bayesian-workflow:setup command installs CmdStan via python -m cmdstanpy.install_cmdstan.

Install

From the marketplace (the repository is also its own marketplace):

/plugin marketplace add sunxd3/bayesian-statistician-plugin
/plugin install bayesian-workflow@sunxd3-plugins

For local development, clone the repository and load it with --plugin-dir:

git clone https://github.com/sunxd3/bayesian-statistician-plugin.git
claude --plugin-dir ./bayesian-statistician-plugin

Run /reload-plugins after editing a locally loaded plugin.

Usage

Bootstrap the Python environment once per project (optional — the orchestrator will do it on first run if you skip):

> /bayesian-workflow:setup

Then run the workflow with your data path or analysis goal:

> /bayesian-workflow:run Analyze data/sales.csv and build a Bayesian model

The orchestrator drives the workflow through four phases, delegating to subagents and writing all results into a predictable folder structure (eda/, design/, experiments/, final_report.html, log.md).

Individual phases can also be run standalone — for example, EDA alone:

> /bayesian-workflow:eda data/sales.csv

What's inside

Orchestration skill — orchestration holds the full workflow protocol (phases, validation-round semantics, canonical file structure, dispatch logic). Loaded by the /bayesian-workflow:run command.

Bundled workflow script — workflows/validate-experiments.js runs one round of the Phase 3 validation pipeline as a deterministic Workflow-tool script: stage sequencing, the FIX-refine budget, MCMC concurrency throttling, and verdict-vs-numbers cross-checks are enforced in code, while the orchestrator keeps every judgment call (critique-driven exploration, new structural questions, model selection) between rounds. On harnesses without the Workflow tool, the orchestration skill falls back to a manual task-pool protocol. Stage agents write a machine-readable status.json completion record, so interrupted rounds resume cheaply — completed stages are verified in seconds instead of re-fitted.

Subagents (11) — eda-analyst, analysis-planner, model-designer, prior-predictive-checker, fake-data-checker, model-fitter, posterior-predictive-checker, critique, model-refiner, model-selector, report-writer.

Commands:

• /bayesian-workflow:setup — bootstraps the Python environment (copies shared_utils, creates pyproject.toml, runs uv sync and cmdstanpy.install_cmdstan).
• /bayesian-workflow:run [data-path-or-analysis-goal] — end-to-end pipeline. Loads the orchestration skill and drives all four phases.
• /bayesian-workflow:eda <data_path> [output_dir] [--focus=<area>] — run EDA on a dataset standalone, without the full workflow pipeline. Wraps the eda-analyst subagent.

Modeling skills (13) — validation-protocol, python-environment, stan (with references/ode.md and references/horseshoe.md for ODE-based dynamics and sparse regression), generative-model-design (lean SKILL.md index + references/ for spec sections, design principles, and the resolution-sequence pattern), fake-data-simulation (with references/single-draw.md, references/sbc.md, references/decision.md), convergence-diagnostics, inferencedata-handling, visual-predictive-checks, bayesian-model-diagnostics, bayesian-model-selection, model-critique (SKILL.md index + references/ for statistical, domain, and framework assessment plus verdict templates), eda (with references/process/ for EDA procedures and references/tests/ for a diagnostic test library by data shape), artifact-guidelines (with references/html-report.md, references/markdown-report.md, and references/final-report.md for the Phase-4 narrative structure). Subagents load the skills relevant to their role; skills are not user-invocable directly — use the commands above.

Bundled library — shared_utils, a Python package with a fit-and-summarize pipeline, convergence diagnostics, LOO, and ArviZ helpers. The setup command copies it into the working project as a path dependency.

How the workflow runs

1. Data understanding (eda/) — eda-analyst explores the data and surfaces competing structural hypotheses about the data-generating process.
2. Model design (design/) — analysis-planner frames the analysis (purpose, validation, domain, structural questions) and the shared baseline; parallel model-designer instances then turn the structural questions into an experiment plan.
3. Model development (experiments/) — each experiment flows through prior-predictive-checker → fake-data-checker → model-fitter → posterior-predictive-checker → critique, with model-refiner and model-selector driving iteration until questions are resolved.
4. Reporting (final_report.html) — report-writer produces a report organized around what was learned about the data-generating process.

Optional settings

Add to your .claude/settings.json:

{
  "env": {
    "CLAUDE_BASH_MAINTAIN_PROJECT_WORKING_DIR": "1"
  }
}

• CLAUDE_BASH_MAINTAIN_PROJECT_WORKING_DIR — resets the working directory to the project root after each bash command, which keeps the canonical folder structure consistent across subagents.

The subagents inherit your main session model, so running Claude Code on an Opus model gives the whole workflow Opus-level quality.

License

MIT — see LICENSE.