CCA Study Tips for Domain 2: Tool Design & MCP Mastery

Study Domain 2 (Tool Design & MCP Integration, 18% of the CCA-F) by reading the 21 articles in task order over three sessions, memorising a short list of file locations and field names, and then practising the four judgement calls the exam keeps returning to: how to describe a tool, how to fail, how many tools, and where a server lives. This post lays out that plan.

For what the domain covers and why it is weighted at 18%, read the companion CCA Domain 2 Explained. This post is only about how to study it.

What order should I read the Domain 2 articles in?

The Domain 2 study guide page is in task order and that order works. Suggested sessions:

Session	Articles	Goal
1	Foundations MCP primer (F3.1 to F3.5) if new to MCP; then Task 2.1: 30% Misrouting with Minimal Descriptions, Data-Driven Description Fixing, Removing Keyword Rules Improved Routing, Three Fields, All Critical	Write a five-element tool description from memory
2	Task 2.2: Structured Errors: 78-95% Recovery vs 15%, error classification, the uniform-error anti-pattern, “Your Order Doesn’t Exist” — But It Does, Retry Locally First	Draw the access-failure vs valid-empty table without notes
3	Tasks 2.3 to 2.5: the tool count curve, tool_choice: auto, any, tool, none, least privilege, MCP server scope, env-var expansion, resources vs tools, auto-discovery, community vs custom, built-in tool responsibilities and Edit fallback	Recite the scope table and the built-in tool one-liners

Sessions are short. If you have hands-on time, spend it after session 3 wiring one community MCP server into a project via .mcp.json with a ${TOKEN} reference, and confirming a teammate (or a fresh clone) gets it automatically.

What should I memorise, and what should I reason about?

Memorise:

• The five description elements: purpose, input format, output format, example use cases, boundary conditions (“do NOT use for X, use Y”).
• isError is a boolean on the tool result; structured detail (category, isRetryable, customer message, suggested action) travels alongside it.
• Access failure means isError: true; a valid empty result means isError: false. Both can look like “nothing found” to a naive handler.
• The tool count curve direction: near-perfect at three to five tools, steep decline past eight, roughly half at eighteen.
• The four tool_choice modes and their guarantees.
• Project scope is .mcp.json at the repo root, committed; user scope is ~/.claude.json, personal. Secrets as ${ENV_VAR} (${ENV_VAR}: Keep Secrets Out of Version Control). CLAUDE.md configures nothing.
• MCP uses inputSchema; the Claude API uses input_schema (The camelCase Trap).
• Built-in tools: Grep searches content, Glob searches paths, Read before Edit, Read plus Write when Edit’s exact match fails (When Edit Fails: Read + Write as Fallback).

Reason about:

• Should this generic tool be split? Yes when one tool serves several operations with different parameter needs and parameter errors are climbing.
• Which error pair is causing the misrouting? Fix the specific pair, not everything.
• Retry locally or propagate? Retry transient errors a bounded number of times inside the tool; propagate business and permission errors with context.
• Where do the fifteen tools go? Into role-scoped specialists of four or five each, with a coordinator (Least Privilege: Each Agent Gets Only Its Role’s Tools).
• Resource or tool? Resources for browsable content the model reads without a call; tools for actions (MCP Resources: Browse Content Without Tool Calls).
• Community server or custom? Community for standard services, custom for proprietary systems (Use Community Servers for Standard Integrations).

The percentages in article titles illustrate direction; the exam does not ask you to reproduce them.

What practice cadence works for Domain 2?

• After each session, do 15 Domain 2 practice questions. Before revealing, classify the question: description, error, count, scope, or built-in tool. Domain 2 questions almost always sit cleanly in one of those five buckets, and naming the bucket usually eliminates two options.
• After session 3, run a 30-question block. Domain 2 is where most people first see accuracy above 80%, which is a useful confidence checkpoint early in a study plan.
• Review mistakes weekly. In this domain, mistakes tend to be either “picked the prompt fix over the structural fix” or “confused two similar file locations”. Both are quick to correct.
• Final week: the free CCA-F mock exam will show roughly 11 Domain 2 questions among 60 (confirm current weights on the official Anthropic / Pearson VUE page before you register).

Self-check questions before moving on

1. Two extraction tools differ by one word in their descriptions. What is the symptom, and what is the fix?
2. A tool handler catches a timeout. What isError value does it return, and what else should the payload include?
3. An agent retried a 403 five times. Which field was missing from the error?
4. Your agent has 12 tools and misroutes a third of calls. What is the structural fix?
5. You need the model to always call some tool but you have three schemas. Which tool_choice mode?
6. A new hire clones the repo and the Jira server is missing. Which file was it in, and which should it be in?
7. Where do you put “use the GitHub server for issues” so it takes effect? (Trick: that sentence configures nothing.)
8. Which built-in tool finds every file that imports a module, and which finds files named *.test.ts? (Grep Each Export to Find All Cross-Module References is the walkthrough.)

Mistakes worth reviewing twice

• Choosing “add a note to the system prompt” when a description or error field is the actual fix.
• Returning isError: true with a generic message and thinking the problem is solved.
• Adding tools “just in case” and expecting selection accuracy to hold.
• Hardcoding a token in .mcp.json because it is “just for the team”.
• Writing a custom Slack or GitHub server when a maintained one exists.
• Reaching for Bash to search when Grep or Glob is the single-job tool (Grep Entry Points, Read Key Files, Trace Imports).

Next step

Do session 1 today from the Domain 2 study guide page, and aim to finish the Domain 2 practice set within the week. Because this domain converges fast, it is a good early win in the 30-day study schedule. When your accuracy holds, take the free CCA-F mock exam to see it in proportion with the other domains.