pymc-labs

Guide a human through creating a dlab decision-pack by asking questions and then calling generate_dpack(). Use this skill whenever the user wants to create, set up, or scaffold a new decision-pack, agent environment, or Docker-sandboxed config for dlab — even if they don't use the word "decision-pack" explicitly. Trigger on phrases like "set up a new agent", "create an environment for X", "I want to run opencode for Y", "scaffold a project", or "make a new config".

How to create a dlab decision-pack directory using generate_dpack() from Python code

Complete reference for decision-lab (dlab). Use when the user asks about creating decision-packs, designing data science agents, running sessions, analyzing results, or anything related to dlab CLI, agent architecture, parallel subagents, or decision-pack configuration. Covers the full workflow from scaffolding to analysis.

Methodology for probabilistic forecasting of when and whether a future event will occur. Covers Bayesian survival models, reference class reasoning, driver threshold models, leading indicator models, scenario decomposition, and causal mechanism models. Use for any question of the form "When will X happen?" or "What is the probability that Y occurs by date Z?"

Design agent system prompts, parallel architectures, and methodological guardrails for data science decision-packs. Use when creating orchestrator, subagent, or parallel agent systems for analytical workflows. Covers anti-fabrication rules, epistemic humility, when to stop, conflict detection, uncertainty reporting, retry protocols, prompt design principles, and the decision-lab runtime mechanics.

Navigate and analyze completed dlab session directories. Use when pointed at a work directory to understand what happened during a run — explore logs, outputs, parallel agent results, and the skills/prompts that shaped the analysis.

Visual language and UX patterns for Textual TUI applications in dlab

Expert on PyMC data management including pm.Data and pm.Minibatch for handling datasets, updating data containers, and mini-batch training. Use for data container errors or dataset handling issues.

Review CausalPy pull requests end-to-end by classifying PR type, checking branch freshness, mergeability, remote CI, correctness, security, tests, docs, and maintainer concerns. Use when asked to review a PR, assess a branch before merge, summarize PR risks, or request changes.

Turn issues into PRs, handle commits, and run prek checks consistently.

Create, evaluate, and triage GitHub issues for CausalPy. Use when filing a bug, proposing an enhancement, analyzing existing issues, or splitting large work into parent-child sub-issues.

Bring a pull request to green by syncing with main, resolving conflicts safely, and fixing failing checks with CausalPy conventions.

Choose the appropriate CausalPy experiment class from a causal or impact question, data structure, treatment assignment, and identification assumptions. Use before writing analysis code when the method is not yet settled, including plain-English questions about whether a campaign, policy, or intervention worked.

Explore unfamiliar APIs, libraries, or implementation behavior with minimal reproducible examples and documented findings. Use when implementation details are unclear and can be resolved by reading docs, inspecting code, and running focused experiments.

Detect, configure, and use a conda-compatible tool. Use before tasks that need the project environment, such as importing project code, running tests, building docs, or invoking repo tooling.

Perform structured research and turn findings into an implementation plan.

Media Mix Modeling with PyMC-Marketing. Use when building MMMs, specifying adstock/saturation transformations, setting priors, fitting multidimensional (geo-level) models, computing channel contributions, ROAS, running budget optimization, calibrating with lift tests, or performing sensitivity analysis. Covers the MMM class, GeometricAdstock, LogisticSaturation, BudgetOptimizerWrapper, and ArviZ diagnostics for marketing models.

PyMC-Marketing coding conventions, preferred implementations, and style guidelines.

Create git commits for changes made during the session.

Create an implementation based on a plan.

Create a plan based on document research through an interactive, iterative process.

Structure a research based on the user request. Identify what must change in order to complete the task.

Learn how to work in the current folder (repository)

Create git commits for changes made during the session.

Load when the user is comparing Bayesian models, computing LOO-CV / ELPD, calling arviz_stats.loo or arviz_stats.compare, doing model stacking/averaging, or computing Bayes factors. Covers the ArviZ 1.1 LOO/ELPD/stacking APIs exclusively (no waic). Triggers include: model comparison, LOO, ELPD, compare, loo_expectations, loo_metrics, loo_r2, Pareto k, stacking, Bayes factor, cross-validation, predictive accuracy, information criterion.

Load when the user is choosing priors, running prior predictive checks, calling find_constrained_prior, using PreliZ, or otherwise eliciting domain knowledge into a Bayesian model. Covers weakly informative priors, constrained priors, sensitivity analysis, and elicitation workflows. Triggers include: prior selection, elicitation, find_constrained_prior, PreliZ, prior predictive, expert/informative priors, weakly informative priors, constrained priors.

Load when the user is working with pymc-extras (pmx) features: splines / BSplineBasis, distributional regression / GAMLSS, R2D2M2CP or horseshoe priors, discrete variable marginalization, or Laplace approximation via fit_laplace. Triggers include: pymc_extras, pymc-extras, pmx, splines, BSplineBasis, distributional regression, GAMLSS, R2D2, horseshoe (regularized/Finnish), marginalize, fit_laplace, penalized splines.

Load whenever the user is working on code that imports pymc, pytensor, or arviz, or asks about Bayesian modeling, MCMC, priors, posteriors, sampling, or model diagnostics. Covers PyMC 6+, PyTensor 3+, ArviZ 1.1+ (DataTree API), pymc-bart, pymc-extras, nutpie, and JAX/NumPyro backends. Use for building probabilistic models, specifying priors, running MCMC, diagnosing convergence, or comparing models. Triggers include: Bayesian inference, posterior sampling, hierarchical/multilevel models, GLMs, time series, Gaussian processes, HSGP, BART, mixture models, prior/posterior predictive checks, MCMC diagnostics, LOO-CV, model comparison, causal inference with do/observe, and any PyTensor Op or graph work.

Load when writing or modifying pytest tests that touch pymc.Model, pm.sample, or any PyMC model code. Covers pymc.testing.mock_sample, pytest fixtures for Bayesian models, and the distinction between fast structure-only tests (mocking) and slow posterior inference tests. Triggers include: testing PyMC, pytest with pymc, unit tests for Bayesian models, mock sampling, test fixtures for models, CI/CD for PyMC.

ALWAYS use when: creating/editing marimo notebooks, working with any .py file containing @app.cell decorators, building reactive Python notebooks, doing exploratory data analysis in notebook form, converting Jupyter (.ipynb) to marimo, or when user mentions "marimo", "reactive notebook", or asks for an interactive Python notebook. Covers marimo CLI (edit, run, convert, export), UI components (mo.ui.*), layout functions, SQL integration, caching, state management, and wigglystuff widgets. If a task involves notebooks and Python, invoke this skill first.

A skill

A skill

Guide for using the dhub CLI — the AI skill manager for data science agents. Covers authentication, publishing, installing, running skills, managing API keys, eval reports, and troubleshooting. Use when users ask about dhub commands, skill publishing workflows, or need help with the Decision Hub CLI.

A skill

Guide for creating effective skills for Claude Code agents. Covers skill design, implementation, validation, packaging, and optionally runtime environments and automated evaluations for Decision Hub publishing. Use when users want to create, improve, or package a skill.

Bayesian statistical modeling with PyMC v5+. Use when building probabilistic models, specifying priors, running MCMC inference, diagnosing convergence, or comparing models. Covers PyMC, ArviZ, pymc-bart, pymc-extras, nutpie, and JAX/NumPyro backends. Triggers on tasks involving: Bayesian inference, posterior sampling, hierarchical/multilevel models, GLMs, time series, Gaussian processes, BART, mixture models, prior/posterior predictive checks, MCMC diagnostics, LOO-CV, WAIC, model comparison, or causal inference with do/observe.

ALWAYS use when: creating/editing marimo notebooks, working with any .py file containing @app.cell decorators, building reactive Python notebooks, doing exploratory data analysis in notebook form, converting Jupyter (.ipynb) to marimo, or when user mentions "marimo", "reactive notebook", or asks for an interactive Python notebook. Covers marimo CLI (edit, run, convert, export), UI components (mo.ui.*), layout functions, SQL integration, caching, state management, and wigglystuff widgets. If a task involves notebooks and Python, invoke this skill first.

A clear description of what this skill does and when to use it. Include specific trigger keywords and task types that should activate this skill. Be specific about capabilities so the AI can decide when to load it. Maximum 1024 characters.

Load whenever the user is working on code that imports pymc, pytensor, or arviz, or asks about Bayesian modeling, MCMC, priors, posteriors, sampling, or model diagnostics. Covers PyMC 6+, PyTensor 3+, ArviZ 1.0+ (DataTree API), pymc-bart, pymc-extras, nutpie, and JAX/NumPyro backends. Use for building probabilistic models, specifying priors, running MCMC, diagnosing convergence, or comparing models. Triggers include: Bayesian inference, posterior sampling, hierarchical/multilevel models, GLMs, time series, Gaussian processes, HSGP, BART, mixture models, prior/posterior predictive checks, MCMC diagnostics, LOO-CV, model comparison, causal inference with do/observe, and any PyTensor Op or graph work.

Load when the user is comparing Bayesian models, computing LOO-CV / ELPD, calling az.loo or az.compare, doing model stacking/averaging, or computing Bayes factors. Covers the ArviZ 1.0 LOO/ELPD/stacking APIs exclusively (no waic). Triggers include: model comparison, LOO, ELPD, az.compare, az.loo, loo_expectations, loo_metrics, loo_r2, Pareto k, stacking, Bayes factor, cross-validation, predictive accuracy, information criterion.

Load when the user is choosing priors, running prior predictive checks, calling find_constrained_prior, using PreliZ, or otherwise eliciting domain knowledge into a Bayesian model. Covers weakly informative priors, constrained priors, sensitivity analysis, and elicitation workflows. Triggers include: prior selection, elicitation, find_constrained_prior, PreliZ, prior predictive, expert/informative priors, weakly informative priors, constrained priors.

Mint short-lived CLI access tokens via the daimon MCP server's get_cli_token tool.

Publish a marimo notebook as a PERMANENT, shareable blog via the daimon MCP server. You mint a one-time upload URL with create_blog_upload_url and curl your notebook file to it — the source never goes through a tool argument. The blog is a live, read-only marimo app (interactive widgets work, code hidden), hosted indefinitely. Use when someone asks to turn a notebook/analysis into a blog post, publish it, or make it shareable.

Publish a marimo notebook for the user via the daimon MCP server's create_notebook_upload_url tool — you mint a one-time upload URL, get the notebook's .py into a sandbox file, and curl -X PUT --data-binary it to the URL (the source never goes through a tool argument, which truncates large notebooks). The curl response returns a slug-as-secret URL the user can open.

Bayesian path analysis (observed-variable SEM) in PyMC. Compiles a lavaan-inspired formula DSL into a generative PyMC model, then layers introspection, identification diagnostics, the `do()` operator, and causal estimands (ATE/CATE/ATT/ATU/prob) on top. Use when the user asks to specify, fit, or query a Bayesian structural causal model; estimate average treatment effects via g-computation; check identification with adjustment sets or the front-door criterion; or simulate panel/longitudinal counterfactuals.