{
  "lesson": "22-skills-and-agent-sdks",
  "title": "Agent Skills: Portable Contract and Runtime Boundary",
  "questions": [
    {
      "stage": "pre",
      "question": "A team has a reusable review method and an existing API that supplies pull request data. Which design assigns each concern to the right primitive according to the runtime policy?",
      "options": [
        "Put the review method in an Agent Skill and expose the API through an MCP tool",
        "Replace the API with a longer user prompt",
        "Put both the method and the API implementation in AGENTS.md",
        "Use a lifecycle hook as the only source of review instructions"
      ],
      "correct": 0,
      "explanation": "A skill carries reusable procedural knowledge, while MCP exposes an external capability. They can compose. AGENTS.md is better for repository-wide defaults, and a hook responds to lifecycle events rather than replacing either layer."
    },
    {
      "stage": "check",
      "question": "Why should a validator require the SKILL.md name to match its containing directory?",
      "options": [
        "It guarantees every host supports the same invocation syntax",
        "It gives discovery and packaging one stable identity instead of two conflicting names",
        "It makes the Markdown heading render correctly",
        "It allows the skill to bypass host permissions"
      ],
      "correct": 1,
      "explanation": "Discovery tools frequently index a skill by both folder and frontmatter. Matching them prevents ambiguous catalog entries and broken installation paths. It does not grant authority or standardize a host's invocation UI."
    },
    {
      "stage": "check",
      "question": "A SKILL.md contains a field that only one runtime understands. What is the portable interpretation?",
      "options": [
        "Move the field into the skill name",
        "Treat it as a host extension and validate it only under an explicit host policy",
        "The field silently becomes part of the core standard",
        "Every runtime must implement it before loading the skill"
      ],
      "correct": 1,
      "explanation": "Portable core metadata and host runtime extensions are separate contracts. An adapter may recognize an extension, but a core validator should not claim that the field has universal semantics."
    },
    {
      "stage": "check",
      "question": "Which primitive best represents a check that must run after every tool call?",
      "options": [
        "An MCP resource",
        "A one-shot prompt",
        "A repository glossary",
        "A lifecycle hook"
      ],
      "correct": 3,
      "explanation": "The defining requirement is event timing, so a lifecycle hook is the right mechanism. A skill may explain a workflow and MCP may expose a capability, but neither inherently fires after each tool call."
    },
    {
      "stage": "post",
      "question": "When is AGENTS.md a better home than an Agent Skill?",
      "options": [
        "When an external service needs an authenticated API call",
        "When a script needs to be executed outside the workspace jail",
        "When conventions should apply throughout work in a repository",
        "When instructions should load only for a rare specialist task"
      ],
      "correct": 2,
      "explanation": "AGENTS.md communicates durable repository context such as build commands and coding conventions. Task-specific procedures fit skills, external operations fit tools, and no instruction file should be used to escape a sandbox."
    },
    {
      "stage": "post",
      "question": "What does a valid Agent Skill grant by itself?",
      "options": [
        "A universal slash command in every agent host",
        "Discoverable procedural instructions, subject to the host runtime's policies",
        "Permission to execute every bundled script",
        "Automatic network access"
      ],
      "correct": 1,
      "explanation": "The portable artifact supplies metadata and instructions. Discovery, invocation UI, tool access, approvals, and script execution remain runtime responsibilities, so validity is not equivalent to authority."
    }
  ]
}