skills.re

What It Is (confidence: medium)

skills.re (“The Registry of Agent Skills”) is an open catalog for distributing procedural knowledge to AI agents. Its core framing: “A skill is an instruction, not an install. Not a package. A practice.” It works with Claude Code, Codex, Gemini CLI, Open Code, Cursor, Windsurf, GitHub Copilot, CodeBuddy, Antigravity, Amp, Trae, Goose, Roo Code, Pi Agent, OpenClaw, Hermes, and more.

Mechanics (confidence: medium)

• CLI install: skills-re install code-review --agent claude → Installed code-review@2.4.1 to .claude/skills/code-review; skills-re sync --agent claude syncs skills into CLAUDE.md. One command, no wiring, no build step.
• Runtime injection: skills mount directly into the agent’s system prompt at runtime — zero glue code on the user side.
• Authoring via URL: a SKILL.md in a public GitHub repo is all it takes; paste the URL at /submit and it’s live “in under a minute.”
• Public review loop: every version is diffable against the last; evals run on every push; reviews are public, attributed, and contestable.
• Sandbox-first operator safety: run any skill against your own fixtures before it handles real work.

Scale and Curation (confidence: medium)

Indexed categories and sizes at capture: Code Frameworks (1,851), Analysis & Insights (1,379), Tools & Platforms (1,188), Process & Methodology (765), Design & Creative (763), Operations & Automation (690), Communication & Strategy (437), Domain Expertise (307) — ~7,380 skills. Editorial layer includes a weekly newsletter (“Editor’s Dispatch”), a curated “Skill of the Week,” and “Editors’ Picks.” Community skill pages show install counts and stars (e.g., clean-code v1.0.0: 53 installs / 164 stars).

Implications: skills.re is the distribution layer of the skill ecosystem — it makes a SKILL.md in a public repo into an installable, versioned, audited artifact for any agent runtime. That turns the authoring guidance of skill-authoring-best-practices into a market: a well-written SKILL.md becomes discoverable and installable across agents rather than tied to one vendor. Its “practice, not package” framing also sharpens the distinction from code-package managers — the unit shipped is instruction text, not compiled artifacts.

Open Questions

• The “skill of the week” and install/star counts are the only quality signals visible at capture; whether they reflect real usage or registry-internal optics is unverifiable from the snapshot.
• Where skills.re routes on agent (--agent claude), OpenAgentSkill routes on task — the two registries are complementary in design but the boundary is untested in practice.
• CLI and sync semantics for agents outside Claude Code (Codex, Gemini) are implied but not shown in the captured page.

Related

openagentskill — the task-to-skill registry counterpart; routes on intent, not agent skills-directory-gate — the runtime pattern skills.re skills mount into skill-authoring-best-practices — the authoring rules the registry distributes and diff-tests skill-reducer — measured evidence on the compressible fat in SKILL.md descriptions ponytail — a real skill (portable across agents) that could be distributed through a registry like this

Sources

• raw/external/skills-re-cc5c6e1d.md