skill vs agents vs claude — which do you need?

Three agent instruction formats compared

Three different conventions for instructing AI agents, each with a distinct purpose and scope.

agents.md website

Quick Answer

  • AGENTS.md — General-purpose instructions for any AI coding agent working on your project (universal convention)

  • SKILL.md — Reusable workflow knowledge for Claude Code (Anthropic-specific format)

  • .claud files — Individual skill definitions for Claude Code (YAML frontmatter + markdown)

In short: AGENTS.md is project-level guidance for agents, SKILL.md is workflow knowledge for Claude, and .claud files are individual Claude skills.

Key Differences at a Glance

Use SKILL.md when:

  • Building a reusable workflow for Claude Code specifically

  • Teaching Claude a repeatable process (code review, testing, documentation generation)

  • Pairing workflow knowledge with MCP integrations (e.g., Sentry + GitHub)

  • You want Claude to consistently follow a specific approach

Example use case: A testing workflow that always checks for edge cases, writes tests in a specific style, and uses your team’s test utilities.

Use .claud files when:

  • Packaging an individual skill for distribution

  • Creating a shareable, installable workflow for Claude Code

  • Building a library of discrete skills users can pick and choose from

  • Publishing to a skill repository or marketplace

Example use case: A “React Component Generator” skill that can be installed and used across multiple projects.

Also Today:

AI Model Theft at Industrial Scale

Anthropic identified three Chinese labs (DeepSeek, Moonshot, MiniMax) conducting massive distillation attacks — 16M exchanges via 24K fraudulent accounts to steal Claude’s capabilities. Stripped-out safeguards create national security risks (bioweapons, cyber ops, surveillance). Attacks undermine export controls by bypassing chip restrictions through theft rather than innovation.

Autonomous Testing for agent-browser

Chris Tate (@ctatedev, Dev @ Vercel) released an autonomous dogfooding skill for agent-browser that explores your app like a user — clicks buttons, fills forms, tests edge cases, checks console, captures repro videos and screenshots, outputs structured report with severity ratings. No test scripts, no manual QA.

Edit this page