CCAR-F (CCA-F) Readiness Check: 12 Free Practice Questions (2026)

1. Your agent loop keeps calling the same tool with the same failing input, burning tokens until you kill the process. What's the architectural fix?

Increase max_tokens so the model has more room to reason its way past the failing call on its own, without any harness changes Return the error in the tool_result and enforce a bounded loop with an exit condition Set temperature to 0 so the model stops varying its approach Switch to a larger model tier

2. A response comes back with stop_reason 'tool_use'. What should your harness do next?

Treat the text so far as the final answer and return it to the user Retry the same request, since the model didn't finish Run the tool, send back a tool_result block, and continue Close the session and surface an error — a tool_use stop means the model cannot complete the request with the information it has

3. A coordinator agent spawns a subagent to audit one module of a codebase. What should the coordinator pass it?

The full conversation history plus all prior tool results, so the subagent starts with everything the coordinator knew Nothing — subagents automatically share the parent's memory The coordinator's own system prompt, unchanged A minimal task brief containing only the context that subtask needs

4. Which of these tasks is a poor fit for a multi-agent architecture?

A linear extraction task that a single well-scoped prompt already handles reliably Reviewing ten independent modules in parallel Adversarially verifying findings produced by another agent A research sweep whose material exceeds one context window

5. Your team has org-wide Claude Code standards, but one legacy repo needs different build commands. Where does each piece of guidance live?

Both in that repo's README so humans see them too Everything in each engineer's ~/.claude/CLAUDE.md Org standards at the shared CLAUDE.md level; the repo's own CLAUDE.md carries the override Skip configuration files entirely and restate the full ruleset at the top of every prompt so guidance is never stale

6. You've built a slash command your whole team should be able to run in Claude Code. Where does it go?

~/.claude/commands on your machine The project's .claude/commands directory, committed to the repo A script entry in package.json An internal wiki page with copy-paste instructions

7. A pipeline extracts structured data from thousands of documents and breaks whenever the model's JSON drifts. The most reliable fix is:

Add 'respond ONLY with valid JSON, no prose, no markdown fences' in capitals at both the start and the end of the prompt Lower the temperature and hope for consistency Add two more few-shot examples of correct output Enforce a schema via tool use / structured output

8. Stable role definitions, non-negotiable rules, and tone guidance belong in:

The system prompt The first user message of each conversation Every user message, repeated for safety Fine-tuning — instructions don't belong in prompts at scale

9. Claude keeps calling your search tool when it should call the lookup tool. Your first fix should be:

Rename the MCP server to something more descriptive Rewrite both tools' descriptions to state when each applies — and when it doesn't Increase the tool call timeout Add a third router tool that inspects each request and decides which of the two underlying tools should handle it

10. The core reason to expose a capability as an MCP server instead of a bespoke integration is:

MCP calls execute faster than direct API calls MCP responses don't count against the context window MCP servers take over authentication and secret management on behalf of every client that connects Any MCP-compatible client can use it without per-client glue code

11. A long-running agent session is approaching the context limit with the task half done. The production-grade move is:

Compact finished work into a state summary and continue from it Raise max_tokens on the next request Switch to a different model mid-conversation to reset the window Keep going — the API automatically truncates the oldest turns and the model degrades gracefully without intervention

12. Your batch pipeline starts receiving 429 rate-limit responses in production. The correct client behavior is:

Retry immediately in a tight loop until the request lands Fail the whole batch and alert on-call Exponential backoff with jitter, honoring any retry-after header Rotate requests across several API keys from different accounts so no single key trips the rate limiter