This domain is 18% of the exam. All five weightings:
Claude does not choose tools by name; it reads their descriptions. This domain is about making those descriptions, boundaries and error responses good enough that the right tool gets called every time.
Write descriptions that route
Two vague descriptions ("gets data", "looks things up") produce constant misrouting. A good description states what the tool returns, the inputs it accepts (with examples), when to use it instead of a similar tool, and the edge cases. When two tools get confused, the first fix is almost always a better description — not a bigger prompt.
Keep each agent's toolset small and focused
Giving one agent a long list of tools degrades selection. Distribute tools by role — a research subagent gets search tools, a synthesis subagent gets none — and restrict each agent to what its job needs.
Tool or Resource?
In MCP, Tools are for actions and Resources are for read-only context that Claude may need to know (a catalog, a schema, documentation). If the capability only exposes data, it should be a Resource, not a Tool. Getting this boundary wrong is one of the most common sources of lost marks.
Structured error responses
"Operation failed" tells Claude nothing. Return what went wrong, the category (for example transient vs validation vs permission), whether it is retryable, and enough detail to decide what to do next. Good error structure is what lets an agent retry a timeout but escalate a permission problem.
Configuration
.mcp.json— project-level server configuration, shared through version control; use${ENV_VAR}expansion for secrets so no credential is committed.~/.claude.json— user-level configuration, handy for personal or experimental servers.
Built-in tools
Know which built-in tool fits which job: Grep to search file contents, Glob to find files by name pattern, Read to inspect a file, Edit for targeted changes and Write to create or replace a file. Using the wrong one (reading a whole file when a search would do) is a typical distractor.
Build to learn it
Build a small MCP server with one Tool (an action) and one Resource (a read-only catalog), connect it to Claude Code, and notice how differently each is used. See the cheat sheet for the configuration files.