Module 09 — Customization: skills, commands, subagents, MCP

Time: 4 hours · Format: 80 min theory + 160 min practice

Theory

1. Why customize

The default agent is generic. Your project is specific. Customization turns the model's general knowledge into your processes.

Four mechanisms:

Mechanism	What it is	When
Skill	a folder with instructions + resources, loaded by description	repeatable process ("how to deploy", "how to write migrations")
Command	a named prompt template (`/review`, `/test`)	frequent requests
Subagent	a separate agent with its own permissions and model	reconnaissance, review, parallel work
MCP	an external server with tools	Jira, DB, browser, internal APIs

2. Skill

A skill is a SKILL.md file with a description of when to apply it:

---
name: deploy
description: How to safely roll out a release to staging and prod
---

# Deploy
1. Check that main is green
2. Build the image: ...
3. ...

The key: description determines when the agent picks up the skill on its own. A bad description = the skill is never used.

3. Command

{
  "commands": {
    "review": {
      "description": "Strict review of the current diff",
      "template": "You are a strict reviewer. Check `git diff` for: ... Problems only."
    }
  }
}

Invocation: /review. Saves typing and standardizes quality.

4. Subagent

A separate session with its own permissions:

{
  "agents": {
    "scout": {
      "description": "Repository reconnaissance without edits",
      "mode": "subagent",
      "permissions": [
        { "action": "edit", "resource": "*", "effect": "deny" },
        { "action": "shell", "resource": "*", "effect": "deny" }
      ]
    }
  }
}

The benefit: it doesn't clutter the main context, and it can't accidentally change code.

5. MCP

A protocol through which the agent gets external tools. Examples: DB access, Jira, browser, internal APIs.

Security rules:

an MCP server is code running with your permissions — only trust verified ones;
prefer OAuth over pasting a key into the config;
restrict: read-only if you only need to look.

6. How not to drown in customization

Start with the problem, not the mechanism:

I repeat the same prompt 5 times a week → Command
There's a 10-step process → Skill
I need to isolate reconnaissance → Subagent
I need data from an external system → MCP

The reverse order ("I'll build an MCP because it's trendy") leaves garbage in the config.

Comprehension checklist

How does a skill differ from a command?
Why is a skill's description more important than its contents?
Why restrict a subagent's permissions?
What is the main risk of an MCP server?

Practice

See practice.md.