CCAR-F cheat sheet: every number and precedence order
The exact names, values and orderings worth having in memory the morning of the exam.
Stop reasons
| Value | Meaning | Your move |
|---|---|---|
tool_use | Claude wants a tool | Execute, append tool_result, loop |
end_turn | Finished naturally | Exit the loop |
max_tokens | Truncated | Do not parse as final; continue or raise the limit |
refusal | Safety stop | Output may not match your schema; handle separately |
stop_sequence | Hit a stop sequence | Handle per your own protocol |
Precedence orders (highest first)
| Settings | MCP server scope | Permission rules |
|---|---|---|
| Managed → command line → project local → shared project → user | local → project → user → plugin → connectors → managed | deny → ask → allow |
CLAUDE.md is the odd one out: files are concatenated root-down, not overridden, with subdirectory files loading on demand.
Numbers to know
| Pass mark | 720 / 1000 scaled · 60 questions · 120 minutes |
|---|---|
| Domain weights | 27 / 20 / 20 / 18 / 15 (D1, D3, D4, D2, D5) |
| CLAUDE.md target | < 200 lines · imports max 4 hops · file > 4 MiB is skipped |
| Cache breakpoints | 4 max · TTL 5 min or 1 h · read 0.1× · write 1.25× / 2× |
| Context editing defaults | trigger 100k input tokens · keep 3 tool uses |
| Batches | 50% off · ≤100,000 requests / 256 MB · ≤24 h · results kept 29 days |
| Strict tools | ≤20 per request · ≤24 optional params · ≤16 anyOf params |
| Subagent limits | depth 3 · concurrency 20 · maxBudgetUsd unset by default |
| Tool token overhead | +286 (auto/none) · +406 (any/tool) |
| Request size caps | 32 MB Messages · 256 MB Batch · 500 MB Files |
| MCP startup timeout | MCP_TIMEOUT, 30 s default |
| Auto memory index | first 200 lines or 25 KB of MEMORY.md |
Config file map
| File | Holds | Committed? |
|---|---|---|
CLAUDE.md / .claude/CLAUDE.md | Project instructions (context) | Yes |
CLAUDE.local.md | Personal project notes | No — gitignore |
.claude/rules/*.md | Topic and path-scoped rules | Yes |
.claude/settings.json | Shared permissions, hooks, env | Yes |
.claude/settings.local.json | Personal overrides | No |
.mcp.json | Project-scoped MCP servers | Yes |
.claude/agents/*.md | Filesystem subagent definitions | Yes |
.claude/commands/*.md | Custom slash commands | Yes |
.claude/skills/<name>/SKILL.md | Skills | Yes |
managed-settings.json | Org policy — cannot be overridden | Deployed via MDM |
Which mechanism for which requirement
| Requirement | Mechanism |
|---|---|
| Guide behaviour, explain conventions | CLAUDE.md / rules |
| Load instructions only for certain files | Path-scoped rules (paths: frontmatter) |
| Package a repeatable procedure | Skill |
| Run something at a fixed lifecycle point, always | Hook |
| Block a tool, command or path outright | permissions.deny (managed settings for org-wide) |
| Isolate context / parallelise / restrict tools | Subagent |
| Orchestrate dozens-plus agents | Dynamic workflow |
| Guarantee output shape | JSON schema output / strict: true |
| Cut cost on bulk async work | Batches API + prompt caching |
| Cut cost/latency on repeated prefixes | Prompt caching |
← Previous: The six exam scenarios | Next: Questions & answers — Domain 1 →
Explore Other Certification Study Guides
Switch to another exam guide to prepare for your next Claude certification:
Claude Certified Associate — Foundations
Entry-level certification for non-coders: prompting, output evaluation, model selection, workflow integration, configuration, governance & responsible use.
Claude Certified Architect — Professional
Advanced enterprise architecture: solution design, system integration, evaluation frameworks, governance, safety & developer enablement.
Claude Certified Developer — Foundations
Build software on Claude: API integration, model selection, tools & MCP servers, agents vs workflows, security, safety & debugging.