CCAR-F Study Guide

CCAR-F facts to memorise: 30 recall cards

Thirty facts that are pure recall — the ones you either know cold or lose a mark on. Cover the right-hand column and work down the list.

PromptWhat you need to know
stop_reason: "tool_use"Claude wants a tool. Execute it, append a matching tool_result, and continue the loop.
stop_reason: "max_tokens"Truncated, not finished. Never parse it as a final answer — JSON in it is likely malformed.
Only channel from parent to subagentThe Agent tool's prompt string. Paths, errors and decisions must be written into it explicitly.
Four benefits of subagentsContext isolation · parallelisation · specialised instructions · tool restriction.
Settings precedenceManaged → command line → project local → shared project → user.
CLAUDE.md loadingConcatenated root-down, not overridden. Subdirectory files load on demand.
Import depth limitFour hops. Imported files still load at launch, so imports organise but do not save context.
Permission rule evaluationdeny beats ask beats allow. Deny cannot be overridden.
Bash(ls *) vs Bash(ls*)With the space: ls, ls -la, not lsof. Without: lsof matches too.
Hook exit code 2Blocks the action; stderr becomes the reason Claude sees. Cannot be overridden by JSON.
Hook exit code 0Success. JSON on stdout is parsed for a permission decision.
MCP tool namingmcp__<server>__<tool>. mcp__* matches every MCP tool anywhere.
MCP server primitivesTools, resources, prompts. (Client side: sampling, elicitation, logging.)
MCP transportsstdio for a local process serving one client; Streamable HTTP for remote, many clients.
Project-scoped MCP file.mcp.json at the repo root — the one you commit. Use ${VAR} for secrets.
strict: trueTool names and inputs conform to their schemas. Different from output_config.format.
output_config.formatJSON Schema constraining Claude's final response. Max 20 strict tools per request.
Unsupported in schemaspattern, allOf, not, if/then/else, dependentSchemas, recursion.
Context editing defaultsTrigger at 100,000 input tokens; keep the last 3 tool uses.
clear_at_leastMinimum cleared per activation, so the cache invalidation is worth its cost.
Cache breakpointsFour maximum, ordered tools → system → messages. Mark the last identical block.
Cache pricingRead 0.1× base. Write 1.25× (5 min TTL) or 2× (1 hour TTL).
Batches API50% off · ≤100,000 requests / 256 MB · most within 1 h, all within 24 h · no streaming.
Retryable HTTP codes429, 500, 504, 529. Not 400, 401, 403, 404, 413.
--bareSkips auto-discovery of hooks, skills, commands, agents, plugins, MCP, memory, CLAUDE.md. Use in CI.
Does not survive /compactThe skill listing, and instructions given only in conversation.
Subagent default limitsDepth 3, concurrency 20, no spend cap unless you set maxBudgetUsd.
disable-model-invocation: trueOnly a human can invoke the skill. Use it for deploys and destructive commands.
Escalation triggerRule-based on observable state — never a model-reported confidence score.
Passing findings between agentsStructured records with source IDs. Prose flattening loses citations for good.

← Previous: Questions & answers — Domain 5  |  Next: Six-week study plan →

Architect CCAR-F Real Exam Questions

Ready to Pass Your CCAR-F Certification?

Practice with authentic exam questions, verified explanations, and timed test simulation to guarantee your success.

Starting from $49.00
Take Practice Exam →