Evergreen Tools
|
← Back to AI Tools

AI Agent Skills

Package your company workflows and hard-won playbooks as reusable skills so a general-purpose agent becomes a specialist on demand: one folder plus a SKILL.md. Anthropic shipped Agent Skills in October 2025 and opened the spec, pairing with MCP as the agent extension stack

Tool Interface

Interactive tool will be available soon

Features

  • ✓ A folder plus a SKILL.md — YAML frontmatter and Markdown instructions — is all a skill needs, with a very low barrier to entry
  • ✓ Progressive disclosure: SKILL.md loads first, and supporting files or scripts are read only when needed, saving context window
  • ✓ Skills compose, so multiple skills stack to complete multi-step work instead of cramming everything into one giant prompt
  • ✓ Portable across Claude apps, Claude Code, the /v1/skills API endpoint and the Claude Agent SDK
  • ✓ Pre-built skills cover Excel, PowerPoint, Word and PDF, and skill-creator can generate the skeleton for you

How to Use

  1. Pick a task you repeat weekly with fairly stable steps — that is the best first candidate for a skill
  2. Generate interactively with skill-creator or hand-write SKILL.md, spelling out trigger conditions, steps, examples and I/O shape
  3. Put long reference material and runnable scripts inside the skill folder, and let progressive disclosure manage context
  4. Regress on real tasks to verify trigger accuracy and output quality, then build the shared team skill library

FAQ

What are Agent Skills?

Agent Skills are a mechanism for packaging instructions, scripts and resources into a folder that a model loads on demand, making it behave more like a specialist. Each skill is self-contained in its own directory and centres on SKILL.md: YAML frontmatter declares the name and description, and the Markdown body carries steps and examples. Anthropic introduced the capability in October 2025 and later opened the Agent Skills spec as a standard, echoing how MCP became the de facto tool-calling standard. Docs at https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview and examples at https://github.com/anthropics/skills

How do Skills compare to MCP, tools and subagents?

MCP and tools answer what an agent can reach — wiring external systems and APIs into the model. Skills answer how the agent knows to do something — teaching it workflows, judgement criteria and output conventions. Subagents are about context isolation and division of labour. The complete picture stacks all three: tools for connectivity, skills for method, subagents for execution in isolated context.

How do I write my first SKILL.md?

The minimum viable version is one directory with one SKILL.md: frontmatter holding the name and description (write the description so it is obvious when the skill should trigger — that drives match rate), and a Markdown body with steps, caveats and examples. If the instructions reference other files, such as a spreadsheet template or a database schema, keep them in the same folder; the model reads them via shell commands only when needed, so they cost context only when used.

How much context do skills consume?

That is exactly what progressive disclosure solves. At startup only the skill name and short description enter context; SKILL.md is read in full only after a user request matches. If the instructions reference further files, those are read only at the step that needs them. Executable scripts are cheaper still — the script code never enters context, and the model receives only the output. So a skill library can be large while per-conversation context stays small.

How should an enterprise manage and distribute skills?

Treat skills as code assets: keep them in version control, review changes, and record an owner and the boundaries of use. For distribution, enable them in Claude app settings, install via the plugin directory for Claude Code, or upload custom skills through the /v1/skills endpoint for API users. Organise by business domain and give extra scrutiny to skills touching credentials or production data — never write secrets into a skill file.

Which platforms support this today?

Per official documentation, Agent Skills covers the Claude apps (available to Pro, Max and Team users once enabled), the Claude Developer Platform and Claude Code, with the /v1/skills endpoint for API users and the Claude Agent SDK combining skills with subagents. Because the spec is open, third-party platforms and tooling are following suit — check the latest official docs when making a selection.