Agentic Customization

Key points

1. 1

   Start with built-in tools. Claude Code already reads, searches, edits files and runs shell commands under permission rules, and the API offers server tools for web search, web fetch and code execution. Do not rebuild these as MCP servers or Skills.

2. 2

   Custom tools are the lightest way to connect Claude to logic inside one application, such as your own library or internal API, with no extra deployment.

3. 3

   Anthropic-schema client tools (bash, text editor, memory, computer use) are trained-in, so Claude calls them more reliably than a custom equivalent, while your code still executes them where you choose.

4. 4

   MCP servers suit external systems and capabilities shared across several clients or teams, with the connection and authentication handled by the server.

5. 5

   Skills package instructions, reference material and scripts for on-demand use. A skill is a folder with a SKILL.md (YAML frontmatter plus instructions) and optional supporting files.

6. 6

   Progressive disclosure: skill descriptions (name and description) are always in context, the SKILL.md body loads when the skill triggers, and bundled files load only when read. Scripts are executed, so only their output uses context.

7. 7

   Write the description to say what the skill does and when to use it, because Claude matches tasks against it. Vague or overlapping descriptions cause wrong or missed triggers. disable-model-invocation: true makes a skill manual-only (/name).

8. 8

   Keep SKILL.md concise (the authoring guidance is under 500 lines), link detailed reference files one level deep, and move deterministic operations into scripts.

9. 9

   CLAUDE.md is loaded every session. Use it for short, always-relevant facts and rules (build commands, conventions, "use pnpm"). Move occasional reference material and procedures into Skills. Files imported with @path still load at launch.

10. 10

    Skill + MCP: MCP provides the connection to a system, and a Skill teaches Claude how to use it well (schema, query conventions, workflows).

11. 11

    Hooks, not Skills or CLAUDE.md, for anything that must happen every time or must be blocked. An instruction is a request; a hook or permission rule is enforcement.

12. 12

    Skills on the Claude API run in the code execution container with no network access and no runtime package installs. Provide live external data through a custom tool or MCP server.

13. 13

    Only install Skills and MCP servers from trusted sources, and audit bundled scripts first. A Skill can direct Claude to run code, and a server can do whatever its process can.

14. 14

    Plugins package Skills, hooks, subagents and MCP servers into one installable unit for reuse across repositories or distribution through a marketplace.

Read the source

• Extend Claude Code
• Extend Claude with skills
• Agent Skills
• Skill authoring best practices
• Server tools