What Are Agent Skills? The Open Standard for AI Coding Tools

If you’ve tried to give Cursor, GitHub Copilot, or Claude Code your team’s coding rules, you’ve run into the same mess: every tool has its own place and format. Cursor uses .cursor/rules and .cursor/skills/, Copilot uses .github/copilot-instructions.md, Claude Code uses CLAUDE.md, and others like Codex, Windsurf, and Gemini CLI each have their own way to inject context. You end up maintaining the same intelligence in five or six places, and when a new tool appears, you start over. Agent Skills are an open standard that fixes that: one way to package team intelligence so any AI coding tool can auto-discover and use it.

The Fragmentation Problem

Today, “tell the AI how we work” is tied to the product. Cursor rules live in the Cursor tree; Copilot instructions live under .github; Claude Code reads a single markdown file. There is no shared schema, no shared discovery mechanism, and no guarantee that what works in one tool will work in another. So teams either duplicate content across tools, pick one and hope, or give up and let each developer paste instructions manually. We wrote more about why that happens and what it costs in why AI tools ignore your team’s patterns. Agent Skills are designed to sit above any single vendor.

What Agent Skills Are

Agent Skills are an open standard (defined at agentskills.io) for packaging team-specific coding intelligence in a single, tool-agnostic format. Think of them as a contract: “here is our conventions document, our anti-patterns, our domain model, and our reviewer preferences.” Any editor or CLI that supports the standard - including Cursor, GitHub Copilot, Claude Code, Codex, Windsurf, and Gemini CLI - can discover and load that package without custom integration. You maintain one source of truth; every compliant tool consumes it. We use this standard in rvue to generate skills automatically from your PR history; you can read the full story in our introducing rvue post.

What an Agent Skill Contains

The spec defines a clear structure so tools know what to expect and how to prioritize.

• Conventions - Naming, file layout, libraries, and style. These can carry confidence or priority so the model knows which rules are hard requirements vs. preferences.
• Anti-patterns - Things the team consistently rejects in review, with severity so security or correctness issues are weighted above style.
• Domain knowledge - Key entities, relationships, and boundaries so the AI understands your product and architecture, not just syntax.
• Reviewer preferences - What reviewers typically look for or comment on, so generated code is more likely to pass review the first time.

That structure is the same whether the skill is hand-written or generated by something like rvue. For implementation details and examples, see our Agent Skills docs.

How Tools Discover and Consume Agent Skills

The standard specifies where to look (e.g. well-known paths and manifest files) and how to interpret the contents. Tools that support it scan the repo (and sometimes parent dirs) for skill definitions, load them at startup or when the workspace changes, and inject the content into the context they send to the model. So Cursor might read from .cursor/skills/, another tool from a different path, but both are reading the same skill format. That’s what makes “write once, use everywhere” possible. For a practical guide to configuring each tool to use team conventions (with or without rvue), check how to make Cursor, Copilot, and Claude Code follow your team’s conventions.

MCP: Real-Time Queries for Richer Context

Static skill files are enough for many use cases, but some tools support the Model Context Protocol (MCP). With MCP, a client can ask a server for specific information at request time - for example, “what are the conventions for this file?” or “what anti-patterns apply to this language?” rvue can generate both a static Agent Skill and an MCP server that answers those questions from the same underlying data. So tools like Cursor, Claude Code, Codex, Windsurf, and Gemini CLI can either load the skill file at startup or query the MCP server when they need deeper, contextual answers. The combination of a portable file format and an optional MCP layer gives teams flexibility without locking them into one pipeline.

Supported Tools and Ecosystems

The goal of the standard is broad adoption. Today, tool support looks like this:

• Cursor auto-discovers skills from .cursor/skills/ and uses them in completions and chat.
• GitHub Copilot can consume the Agent Skills spec so the same skill works in Copilot-powered workflows.
• Claude Code, Codex, Windsurf, and Gemini CLI can use skill files and/or MCP, depending on their implementation.
• Any environment that supports MCP can talk to an rvue-generated (or custom) MCP server for on-demand team intelligence.

As more editors and CLIs adopt the spec or MCP, the same skill you create or generate today will work in more places without extra work. For an overview of how AI fits into modern code review across these tools, see our AI code review guide for 2026.

How rvue Generates Agent Skills From PR History

Writing a comprehensive skill by hand is possible but time-consuming. rvue automates it: you run npx rvue-cli enable, it scans your codebase and fetches PR history from GitHub, runs an AI analysis, and outputs an Agent Skill (and optionally an MCP server) that matches the open standard. Conventions, anti-patterns, and domain knowledge are derived from real code and review comments, so the skill reflects how your team actually works. That output is then consumable by Cursor, Copilot, Claude Code, Codex, Windsurf, Gemini CLI, and any other tool that supports the spec or MCP. We walk through the full flow in introducing rvue and in getting started.

Why Open Standards Matter for AI-Assisted Development

Proprietary config formats lock you into a single vendor. When the next great assistant appears - whether it’s a new Cursor, a new Copilot, or something we haven’t seen yet - you don’t want to rebuild your team’s intelligence from scratch. An open standard like Agent Skills means:

• You maintain one artifact that any compliant tool can use.
• New tools can adopt the spec and immediately benefit from existing skills.
• The ecosystem can improve the spec and tooling without depending on a single company.

That’s why we built rvue on Agent Skills and MCP: we want team intelligence to be portable and durable, no matter which AI coding tool you or your team chooses tomorrow.

Next Steps

To see the spec in practice, read the rvue docs and the Agent Skills section. To generate a skill from your repo in one command, start with getting started and run npx rvue-cli enable. Your team’s conventions, anti-patterns, and domain knowledge can then flow to Cursor, GitHub Copilot, Claude Code, Codex, Windsurf, Gemini CLI, and any tool that supports the open standard.