castorini-skills

castorini-skills is a shared agent skill library for cross-repo workflows in the Castorini research ecosystem.

This repository packages reusable skills that are useful across multiple Castorini repositories. It is intentionally small and curated: each shared skill should cover a repeatable cross-repo workflow or contract that would otherwise be duplicated in several codebases.

What lives here

The shared skills in this repository live under skills/.

Skill	Purpose
anserini-fatjar	Quick Anserini setup and smoke-test workflow via the published fatjar on Maven Central
castorini-cli-reference	Shared CLI contract reference for nuggetizer, ragnarok, rank_llm, and umbrela
castorini-onboard	Development environment setup for one or more Castorini Python repos
castorini-pipeline	End-to-end retrieval, reranking, generation, nuggetization, and judging workflow coordination
castorini-release	PyPI and TestPyPI release workflow guidance for Castorini Python packages
castorini-serve	Local HTTP serving recipes and curl/jq pipelines for Anserini, rank_llm, ragnarok, nuggetizer, and umbrela

Shared Skills vs Repo-Local Skills

Use this repository for skills that apply across repositories.

Use repo-local skills when the workflow is tightly coupled to one repository's code, tests, data layout, or release process. Those skills should live inside the repository that owns the workflow.

Today the shared skills complement repo-local skill sets that already ship with the Castorini repositories. For example:

• nuggetizer ships repo-local quickstart, verification, and evaluation skills
• ragnarok ships repo-local quickstart, verification, and dataset skills
• umbrela ships repo-local quickstart, verification, and evaluation skills

rank_llm also uses the shared onboarding, CLI-reference, pipeline, and release skills in this repository for cross-repo workflows.

Installation

Claude Code skillSources

Add this repository as a remote skill source in Claude Code:

{
  "permissions": {
    "allow": ["Bash(git clone*)"]
  },
  "skillSources": [
    "git@github.com:castorini/castorini-skills.git"
  ]
}

Local clone

Clone the repository locally:

git clone git@github.com:castorini/castorini-skills.git
cd castorini-skills
./scripts/install-skills.sh list
# Default: install into the current workspace root.
./scripts/install-skills.sh add -a claude
./scripts/install-skills.sh add -a claude-code
./scripts/install-skills.sh add -a codex
# Optional: use a supported home-directory location instead.
./scripts/install-skills.sh add --home -a claude
./scripts/install-skills.sh add --home -a claude-code
./scripts/install-skills.sh add --home -a codex
# Or install into any directory (each skill becomes a subfolder there):
./scripts/install-skills.sh add -d path/workspace/skills

The installer discovers skills from each SKILL.md file and copies them into one of three locations:

• -a <agent>: the default agent layout under the current workspace root
• --home -a <agent>: the supported home-directory layout for that agent
• -d <path>: an explicit custom destination

Current home-directory targets are supported for claude-code and codex. Other agents should use the default workspace-root layout or -d. claude is accepted as an alias for claude-code.

For more information, see ./scripts/install-skills.sh help.

Supported agents:

• claude (alias for claude-code)
• claude-code
• codex
• cursor
• gemini-cli
• github-copilot
• windsurf
• cline
• roo
• opencode

Manual symlink fallback

mkdir -p .claude/skills
ln -s /path/to/castorini-skills/skills/anserini-fatjar .claude/skills/anserini-fatjar
ln -s /path/to/castorini-skills/skills/castorini-cli-reference .claude/skills/castorini-cli-reference
ln -s /path/to/castorini-skills/skills/castorini-onboard .claude/skills/castorini-onboard
ln -s /path/to/castorini-skills/skills/castorini-pipeline .claude/skills/castorini-pipeline
ln -s /path/to/castorini-skills/skills/castorini-release .claude/skills/castorini-release
ln -s /path/to/castorini-skills/skills/castorini-serve .claude/skills/castorini-serve

Updating Skills

Skills installed from a clone are copied or linked into an agent-specific directory. They do not update themselves automatically.

To pick up changes:

git pull origin main

Then relink or reinstall the skills you use.

Why Selective Installation Matters

Installed skills add routing context for the agent. A larger installed set gives the model more candidate skills to consider on every request, which can increase trigger overlap and reduce precision.

Prefer installing only the shared skills you actively use in a workspace instead of installing every available skill by default.

Acknowledgements

This skills package was informed by publicly shared work from others building practical skill systems for coding agents.

• Thariq from Anthropic for sharing Lessons from Building Claude Code: How We Use Skills
• Elastic for publishing the open elastic/agent-skills repository