Universal Skills

Version: 1.6.0

Universal Skills is a portable catalog of agent capabilities and dependency-ordered skill workflows. It supports Claude-compatible SKILL.md discovery and the meristem provider ecosystem.

Catalog model

• universal_skills/<domain>/<name>/ contains one atomic skill with one capability.
• universal_skills/<domain>-workflows/<name>/ contains a pure DAG over atomic skills or exact MCP tools; workflows do not embed business logic.
• Platform and package operations belong under agent-packages/agents/<package>/<module>/skills/ and are exposed through the package's scion_sdk.connector entry point.
• assets/ may contain nested templates, but those are package resources rather than canonical catalog entries.

The catalog includes development, infrastructure, finance, research, product, creative, documentation, web, system, operations, content, and data capabilities. The content and data domains include evidence-aware drafting, publication preflight, dataset profiling, deterministic data-quality validation, and data-dictionary authoring.

Connector packages use mcp-builder for SDK scaffolding and shared pipeline contracts.

Catalog quality and roadmap

• Per-skill structural audit
• Category improvement roadmap and workflow designs
• Skill and workflow architecture
• Published documentation
• Contribution and specification workflow

Reproduce the catalog audit locally:

python universal_skills/skills/skill-catalog-auditor/scripts/audit_catalog.py \
  universal_skills --format markdown

The audit reports trigger quality, atomicity signals, workflow dependency problems, unbound components, placeholder content, team-configuration drift, broken local links, and taxonomy consistency. It is a structural lint, so safety and domain correctness still require human review.

Install

Install the package, then deploy discovered skills to supported agent tools:

pip install universal-skills
install-skills --all-detected --symlink

For local development:

git clone https://github.com/Knuckles-Team/universal-skills.git
cd universal-skills
pip install -e '.[test]'
install-skills --all-detected --symlink

The repository also provides install.sh and install.ps1 for package installation and skill deployment. Review targets and configuration changes before using broad installation options.

Use from Python

from pydantic_ai_skills import SkillsToolset

from universal_skills.skill_utilities import (
    get_skill_graph_path,
    get_universal_skills_path,
)

skills = SkillsToolset(
    directories=[get_universal_skills_path(), get_skill_graph_path()]
)

Discovery respects each skill's enable/disable environment flag. Package-owned skills are discovered from installed provider entry points by the universal installer.

Author a skill or workflow

Use the bundled builders rather than copying a legacy entry:

# Atomic skill: creates only explicitly requested resource directories
python universal_skills/skills/skill-builder/scripts/init_skill.py --help

# Pure DAG workflow: resolves atomic skills or explicit MCP tools
python universal_skills/skills/skill-workflow-builder/scripts/build_workflow.py --help

Before opening a change:

python scripts/gen_bumpversion.py
bump2version --config-file .bumpversion.cfg --allow-dirty --no-commit --no-tag patch
python scripts/check_atomicity.py --strict
pytest -q
pre-commit run --all-files

Security boundaries

Audit and discovery are read-only. Skills that publish, send, delete, rotate, restart, deploy, trade, persist data, or otherwise change external state must make that side effect explicit and obtain the authorization appropriate to the target. Network scripts must verify TLS by default, bound requests, redact secrets, and treat retrieved content as untrusted input.

License

MIT. See LICENSE.