Official Exam Facts (CCAR-F Exam Guide)
- Cert page: $125 · four core areas (Code, Agent SDK, API, MCP) · must review Exam Guide + Terms + Cert Policy first.
- Exam Guide: 60 items/120min · multi-choice & multi-response (each states how many to pick) · scaled 720/1000 pass · 12-mo valid · 4 of 6 scenarios · proctored (online and/or test center) · pass/fail + percent-correct by domain.
- Guide itself can change — re-check before booking. ~6mo hands-on experience is a community rule of thumb, not an official figure.
Four Core Areas
- Claude Code — agentic coding CLI.
- Claude Agent SDK — build custom agent apps.
- Claude API — messages, tools, streaming, caching…
- MCP — standard protocol for tools/data/prompts.
5 Domains (official weights)
| Domain | Wt |
| Agentic Architecture & Orchestration | 27% |
| Tool Design & MCP Integration | 18% |
| Claude Code Configuration & Workflows | 20% |
| Prompt Engineering & Structured Output | 20% |
| Context Management & Reliability | 15% |
Scenario Bank (4 of 6 drawn)
- 1. Customer Support Resolution Agent — Agent SDK, MCP tools (get_customer/lookup_order/process_refund/escalate_to_human), 80%+ first-contact resolution.
- 2. Code Generation with Claude Code — CLAUDE.md, permission modes, subagent handoff.
- 3. Multi-Agent Research System — lead agent + subagents, fan-out, synthesis.
- 4. Developer Productivity with Claude — review/docs/refactor automation vs human control.
- 5. Claude Code for CI — non-interactive mode, hooks as gates, rate-limit/cost discipline.
- 6. Structured Data Extraction — tool-schema forced output, field validation.
Agent Loop Anatomy
- Gather context → act → verify → repeat, until stop.
- API mechanics: send messages+tools → model emits tool_use → you execute → return tool_result → repeat until end_turn.
- Always define a stop condition: turns, tokens, or human interrupt.
Prompt vs Agent vs Multi-Agent
| Signal | Use |
| Everything's in the input, one pass | Prompt |
| Needs live data/action, bounded | Tool-using agent |
| Decomposes / parallelizable | Multi-agent |
Default rule: start simplest, escalate only as needed. Every step up = more cost/latency/failure surface.
MCP Primitives
| Primitive | Driven by |
| Tools | Model-controlled |
| Resources | Application-controlled |
| Prompts | User-controlled |
| Sampling | Server-initiated (asks client's model) |
MCP Transports
- stdio — local subprocess, simple, no network. Claude Code default for local servers.
- Streamable HTTP — remote/shared servers, multiple clients, streaming.
Host = app (Claude Code). Client = 1:1 connection inside host. Server = exposes tools/resources/prompts.
Tool Schema Design Rules
- Description is a prompt — vague description → wrong tool / bad args.
- Explicit types, constraints, enums over free text.
- Granularity: not too narrow (tool-selection + context overhead), not too broad (ambiguous args).
- Document idempotency & side effects; gate non-idempotent/destructive tools.
Claude Code Features
| Feature | One-liner |
| CLAUDE.md | project instructions, auto-loaded |
| Subagents | separate context, bounded subtask |
| Hooks | deterministic shell cmds at loop points |
| Skills | on-demand packaged instructions |
| Slash commands | reusable custom workflows |
| MCP integration | Claude Code = MCP host |
Permission Modes
- Interactive (default) — approve each sensitive action.
- Plan mode — research/propose only, no edits until approved.
- Auto-accept edits — edits auto-approved, other sensitive actions still gated.
- Autonomous/bypass — no prompts; sandboxed/CI use only, not a production default.
Prompt Engineering Ladder
- Clear direct instructions
- System prompt for standing context
- Examples (few-shot)
- Chain-of-thought / extended thinking
- Structured sections (XML tags)
- Prefill / constrained (tool-schema) output
Prompt → retrieval → fine-tune, in that cost order. Try cheapest fix first.
Context Budget — What Consumes It
- System prompt + every tool def (costs tokens even unused) + history + retrieved docs + tool results.
- Context rot: quality degrades as window fills, before hitting the hard limit.
- Fix: trim stale tool output, summarize/compact, scope tools to the task.
Prompt Caching Rules of Thumb
- Mark a stable prefix (system prompt, tools, big doc) cacheable → cuts cost+latency on reuse.
- Best fit: long stable content reused across many requests, fast (cache expires after inactivity, ~minutes by default).
- Stable content first, variable content last — cache matches a prefix.
Retrieval vs Long Context
| Long context | Retrieval |
| Fit | small/bounded, reused often | large/changing/huge corpora |
| Cost | caching helps | pay per fetch + infra |
| Complexity | simple | index/retriever needed |
Model Selection
| Tier | Fit |
| Opus | hardest reasoning, highest cost/latency |
| Sonnet | balanced default workhorse |
| Haiku | fast/cheap, high-volume simple tasks |
Mixing tiers within one architecture (cheap for bulk, top tier for the hard step) is expected, not a shortcut.
Streaming vs Batch vs Sync
- Streaming — interactive user waiting; improves perceived latency only.
- Sync — short programmatic call, no one watching it stream.
- Batch — high volume, non-urgent, discounted; trades latency for cost.
Evaluation & Guardrails
- Evals = held-out task set, run on prompt/tool/model change → catch systemic regressions.
- Guardrails = runtime layers: input classify, tool-level constraints, output validation → catch bad individual outcomes.
- Complementary, not substitutes.
Cost Levers
- Model-tier mixing · prompt caching · batch for non-urgent volume · shorten prompts before adding infra.
- Per-task token/turn/dollar budgets; alert on anomalous spend.
Failure Modes & Fixes
| Failure | Fix |
| Looping/thrashing | step/retry budget |
| Context rot | compaction/context editing |
| Tool misuse | better schema/description |
| Ungrounded claims | citations, verification |
| Runaway cost | explicit budgets |
| Unsafe autonomous action | human checkpoint, least privilege |
Keyword → Answer
- "needs live data/action" → tool-using agent
- "decomposes / parallel" → multi-agent
- "repo conventions repeated every session" → CLAUDE.md
- "before this runs it must be approved" → human-in-the-loop / plan mode
- "same tool needed by many apps" → MCP server
- "answers get worse the longer it runs" → context rot → compact
- "huge / changing corpus" → retrieval
- "same big prompt reused often, fast" → prompt caching
- "high volume, no one waiting" → batch
- "guaranteed parseable output" → structured output / tool schema
Top Traps
Tools(model) vs Resources(app) vs Prompts(user) · retrieval vs caching-eligible long context · multi-agent as default overkill · agent with no human checkpoint before irreversible action · caching used to mask a bloated prompt instead of shortening it · assuming the Exam Guide can't change between now and your booking
Ask what the scenario optimizes for — cost / latency / accuracy / autonomy&safety / simplicity — then pick the simplest architecture that satisfies it.