Bayesflow Skills

claude-skills

A monorepo of Claude Code plugins for scientific computing, code review, and workflow automation.

Plugins

bayesflow-skills

Domain-specific skills for BayesFlow 2.x extension packages. Provides patterns and best practices for building BayesFlow extensions.

Skill	Description
bayesflow-simulator	Data-generating processes with @allow_batch_size, RNG discipline, and prior/likelihood/meta functions
keras-ops	Backend-agnostic tensor math using keras.ops.* for loss functions, layers, and forward passes
bayesflow-adapter	Data preprocessing pipelines, three-slot schema, and standardization transforms
bayesflow-testing	Test patterns with conftest.py setup, shape testing, and mocking
bayesflow-validation	Simulation-based calibration (SBC), coverage metrics, and quality gates
bayesflow-memory	GPU memory management, OOM recovery, and gradient checkpointing
bayesflow-packaging	Package structure, __all__ exports, pyproject.toml, and CI configuration

Skills trigger automatically based on task context (e.g., creating a simulator, debugging OOM errors, writing tests). You can also invoke them explicitly via the bayesflow-skills: namespace.

---

code-sentinel

Independent code review via specialized subagents. Reviews uncommitted changes for bugs, security issues, silent errors, and logical inconsistencies. Specialized for R and Python codebases.

Skill	Description
code-sentinel	Orchestrates independent review subagents that audit your diff with no prior context of the implementation

Usage: Ask Claude to "review my changes", "check my code before committing", or "audit this diff". The skill delegates to subagents that have no knowledge of the implementing conversation, ensuring unbiased review.

---

workflow-automation

A Plan/Implement/Ship workflow for structured feature development. Breaks large features into phases, implements them in worktrees, and ships each phase as a PR with CI monitoring.

Skill	Description
plan	Create a structured design document before implementation — scope, decisions, and phased steps
implement	Build from a plan document phase by phase, with cross-session resume via state tracking
ship	Commit, create a PR, monitor CI, and merge a completed phase
implement-ship	Implement a phase and ship it (commit, PR, CI, merge) in one session
implement-ship-all	Autonomously implement and ship all remaining phases from a plan
monitor-pr	Background CI monitoring with auto-merge — checks status every 4 minutes and fixes failures

Usage (after installation): Start with /plan to design a feature, then /implement to build it phase by phase, and /ship to get it merged. For a fully autonomous flow, use /implement-ship-all (or say "autopilot") to run from plan to merged PRs without stopping.

---

scientific-literature

Domain-agnostic literature management for research projects. Covers OpenAlex search and DOI validation, APA 7 citation generation, LaTeX-safe BibTeX export, full-text acquisition, extraction, citation crawling, and lightweight pipeline orchestration.

Skill	Description
lit-search	Search OpenAlex for candidate papers and inspect normalized metadata
lit-validate	Validate a DOI or title against OpenAlex before writing references
lit-cite	Add a normalized citation to _index.json and regenerate references.md
lit-latex	Export references.bib, lint BibTeX formatting, resolve citekeys, check .tex files, and optionally smoke-test LaTeX compilation
lit-acquire	Locate full text via Zotero, OpenAlex OA links, or arXiv
lit-extract	Extract structured markdown from PDFs
lit-crawl	Expand forward and backward citation chains
lit-run	Orchestrate multi-step review pipelines from tracked state
lit-status	Inspect pipeline status without side effects

Usage: Ask Claude to search papers, validate references, generate citations, or maintain a LaTeX bibliography. For BibTeX verification, the plugin now includes offline sample fixtures and scientific-literature/scripts/verify_latex.py.

Installation

Install plugins using Claude Code's /install-plugin command. See the Claude Code plugins documentation for the latest syntax.

# Install from GitHub directly
/install-plugin matthiaskloft/claude-skills/bayesflow
/install-plugin matthiaskloft/claude-skills/code-sentinel
/install-plugin matthiaskloft/claude-skills/workflow-automation
/install-plugin matthiaskloft/claude-skills/scientific-literature

How it works

Once installed, skills activate automatically based on what you ask Claude to do. Each plugin adds its skills to Claude Code's context, and the matching skill triggers when your request fits its description.

You can also invoke any skill explicitly using slash commands (e.g., /plan, /ship) or by referencing the plugin-qualified name (e.g., bayesflow-skills:bayesflow-simulator, workflow-automation:plan).

License

MIT