Claude Code Skills

Agent skills for running HiBayES analyses and making the statistical decisions inside them

Overview

HiBayES ships two agent skills for Claude Code, in .claude/skills/. A skill is a Markdown file the agent loads when the task matches its description — these two give an agent the package’s mechanics and its statistical discipline, so it configures analyses that follow the workflow rather than guessing at YAML.

Skill When it fires
running-hibayes Installing, running, configuring, extending, or debugging an analysis: the five-stage pipeline, the config surface, the CLI, the component registries, watching the TUI, custom components, and the AnalysisState output layout.
hibayes-statistical-workflow Statistical decisions inside that pipeline: stating an estimand, choosing a likelihood, how much pooling, priors and predictive checks, MCMC diagnostics, model comparison, and reporting. Ships a plain-language concepts reference for explaining results to non-Bayesians.

The two cross-reference each other, and an agent doing a real analysis typically uses both.

Installation

In a HiBayES checkout

Nothing to install. The skills live in .claude/skills/ in this repository, so Claude Code picks them up automatically when it runs in a clone:

git clone https://github.com/UKGovernmentBEIS/hibayes.git
cd hibayes
claude

Ask “which model should I use for these judge scores?” or “analyse these eval logs with hibayes” and the relevant skill loads.

In your own project

Analysing eval data in your own repository is the more common case. Install the skills as a plugin:

/plugin marketplace add UKGovernmentBEIS/hibayes
/plugin install hibayes-skills@hibayes

Only the .claude/ directory is fetched, so this does not pull the repository’s example data.

To have updates arrive automatically, add the following to ~/.claude/settings.json (user-wide) or your project’s .claude/settings.json (shared with everyone who opens it) instead of running the commands above:

{
  "extraKnownMarketplaces": {
    "hibayes": {
      "source": { "source": "github", "repo": "UKGovernmentBEIS/hibayes" },
      "autoUpdate": true
    }
  },
  "enabledPlugins": { "hibayes-skills@hibayes": true }
}

enabledPlugins is what installs and enables the plugin at session start; registering the marketplace alone installs nothing. To update manually: /plugin marketplace update hibayes, then /reload-plugins.

Alternatively, copy .claude/skills/running-hibayes/ and .claude/skills/hibayes-statistical-workflow/ into your project’s own .claude/skills/.

What the skills change

  • Verified reference tables. Component names, config keys, CLI flags and defaults are checked against src/hibayes/, so the agent uses hibayes-comm and data_process: rather than plausible-looking alternatives.
  • Unattended runs work. The agent knows to pass --no-tui and set interactive: false on the predictive-check plots, and how to hand you a live TUI in tmux when you want to watch a fit or approve a plot yourself.
  • The workflow is not optional. Prior predictive checks, convergence diagnostics and posterior predictive checks come before conclusions, and a failing diagnostic is reported rather than worked around.
  • Judge analyses start from the paper. Grader and judge questions route to the runnable configs in examples/llm-as-a-judge-usecases/.
  • Decisions are explained. Each statistical choice is stated with its reason in plain language, and the concepts reference supplies fuller explanations when they’re wanted.

Contributing to the skills

The skills are plain Markdown in .claude/skills/, with no build step — edit them and open a pull request. Two conventions: verify claims against src/hibayes/ rather than the docs (registered component names are the defs directly under the @extractor, @process, @model, @checker and @communicate decorators), and bump the “Last reviewed” footer in each SKILL.md when you check it against a new version.