CCAR-F scenarios › Developer Productivity with Claude · Question 32 of 60

Scenario 4 · Developer Productivity with Claude

Your team is building internal developer tools on the Claude Agent SDK. The agent helps engineers find their way around unfamiliar or legacy codebases, generates boilerplate, and automates repetitive chores. It works through the built-in Read, Write, Bash, Grep, and Glob tools, plus connections to Model Context Protocol (MCP) servers.

Tool Design & MCP Integration

Your team connected an internal MCP server whose find_symbol_references tool resolves references through re-exports, aliases, and dependency-injection bindings. Its description reads only "Finds references." Session logs show the agent nearly always uses the built-in Grep instead, and so it misses aliased call sites when engineers ask about legacy services. What is the most effective first step?

Show answer and explanation

Answer: C — Expand the tool's description to explain that it resolves aliases, re-exports, and injected bindings, what it returns, and when to prefer it over Grep.

Tool descriptions are the main signal the model uses to choose a tool. Section 6 specifically calls for enhancing MCP tool descriptions so the agent stops preferring built-in tools like Grep over a more capable MCP tool. (a) is too blunt: Grep is still the right tool for string literals and error messages. (b) adds engineering work in a hook to make up for a weak description and wastes a call on every search. (d) is a broad keyword-style instruction that sends the agent to MCP tools even where they don't fit, and it still leaves the agent with no idea what this tool actually does.