universal-agent-skills

Universal Agent Skills

███████╗██╗  ██╗██╗██╗     ██╗     ███████╗
██╔════╝██║ ██╔╝██║██║     ██║     ██╔════╝
███████╗█████╔╝ ██║██║     ██║     ███████╗
╚════██║██╔═██╗ ██║██║     ██║     ╚════██║
███████║██║  ██╗██║███████╗███████╗███████║
╚══════╝╚═╝  ╚═╝╚═╝╚══════╝╚══════╝╚══════╝

A canonical, version-controlled collection of Agent Skills that can be shared across Codex, Claude Code, OpenCode, GitHub Copilot, and compatible coding agents.

The repository is both:

• a portable skill source under skills/;
• a native Claude Code and Codex plugin at the repository root.

The installers are idempotent, support symbolic links or copies, track what they own, refuse unmanaged conflicts by default, and provide dry-run and uninstall modes.

What is included

.
├── skills/                         # Canonical Agent Skills
│   ├── coding-style/
│   ├── simplify-code/
│   ├── surgical-implementation/
│   └── verify-changes/
├── profiles/default.json           # Desired external plugin/skill stack
├── adapters/agents.tsv             # Agent discovery paths
├── .claude-plugin/                 # Claude plugin + marketplace metadata
├── .codex-plugin/                  # Codex plugin metadata
├── .agents/plugins/marketplace.json
├── install.sh / install.ps1        # Local install, sync, and uninstall
├── bootstrap.sh / bootstrap.ps1    # Clone/update + install
├── scripts/sync_instructions.py    # Safe global instruction merge/remove
└── .github/workflows/              # Validation, tests, and scanners

coding-style tells agents to prefer self-explanatory code and use short, neutral comments only when intent or constraints are not obvious. It explicitly avoids conversational, first-person, AI-like narration.

simplify-code provides the portable equivalent of a focused code-simplifier workflow while preserving behavior. surgical-implementation keeps changes small, makes material assumptions visible, and prevents drive-by refactoring. verify-changes requires focused validation evidence before an agent reports completion.

Supported targets

Agent	Global target	Project target	Native plugin
Codex	~/.agents/skills	.agents/skills	Yes
Claude Code	~/.claude/skills	.claude/skills	Yes
OpenCode	~/.config/opencode/skills	.opencode/skills	Skills directly
GitHub Copilot	~/.agents/skills	.agents/skills	CLI plugins; editor uses skills/instructions
Other Agent Skills clients	~/.agents/skills	.agents/skills	Client-dependent

OpenCode also discovers the Claude-compatible and .agents/skills locations. GitHub Copilot and Codex share the standard .agents/skills target, so the copilot and universal adapter names are aliases of codex and do not create duplicate skills.

See Agent compatibility for the source documentation and discovery behavior.

Install from a clone

git clone https://github.com/26zl/universal-agent-skills.git
cd universal-agent-skills
./bootstrap.sh

The default is a global symbolic-link installation for Codex/GitHub Copilot, Claude Code, and OpenCode. Editing a canonical skill updates every linked agent immediately.

Existing plugins and skills installed through each agent's own manager are never touched, so nothing needs to be uninstalled first: the installer only manages its own entries and stops on unmanaged name conflicts, and the stack reconciler leaves everything outside the profile unchanged.

Use copies when symbolic links are unavailable or undesirable:

./install.sh --mode copy

Install only selected agents or skills:

./install.sh --agents codex,opencode
./install.sh --skill coding-style

Install into the current project instead of the user profile:

./install.sh --scope project

Bootstrap one-liner for macOS and Linux

Pin a release tag or full commit for reproducible installs: