Skip to content

CCAR-F One-Pager: Claude Certified Architect – Foundations

Unofficial summary. Every exam fact on this page comes from 16-certification-tracks/registry.yaml (verified 2026-10-24 against the official exam guide v1.0, effective July 2026) and is summarised in exam-facts.md. The official guide on the Anthropic Partner Academy is the source of truth. Pearson VUE has announced updated guides for 2026-11-17, so re-check before you book. Full track: 16 · CCAR-F.

Exam format

Item Value (from the registry)
Audience Solution architects building production applications with the Agent SDK, Claude Code, MCP and structured output
Recommended experience 6+ months building with the Claude API, Agent SDK, Claude Code and MCP (not mandatory)
Items 60, multiple-choice and multiple-response
Structure 4 scenarios drawn at random from a published bank of 6
Time 120 minutes
Passing Scaled score 720 on a 100–1,000 scale
Fee USD 125 (list price)
Validity 12 months
Delivery Pearson VUE, online proctored or test centre
Eligibility Organisations in the Claude Partner Network

The six scenarios and how to study each are in scenarios.md.

Domains, weights and what to know

D1 · Agentic Architecture & Orchestration — 27% (largest domain)

  • The agent loop: call, run tools, return results, stop on end_turn; cap turns and budget (CCAR-F-1.1).
  • Coordinator and subagents: delegate for parallelism or context isolation, return summaries, not transcripts (CCAR-F-1.2, 1.3).
  • Subagents are spawned with the Agent tool (called Task in the v1.0 guide); pass what they need in the prompt.
  • Enforce order with code, not prose: prerequisite checks and hooks beat "always do X first" (CCAR-F-1.4).
  • Hooks (PreToolUse, PostToolUse) intercept, block or normalise tool calls and results (CCAR-F-1.5).
  • Decompose by independence and context size; keep sessions resumable or fork them (CCAR-F-1.6, 1.7).
  • Study: 10 · The agent loop, 10 · Multi-agent architecture, 09 · Hooks.

D2 · Tool Design & MCP Integration — 18%

  • Tool descriptions are the interface: say what, when, inputs, outputs and limits (CCAR-F-2.1).
  • Return errors as results with isError and enough context to recover; don't hide failures (CCAR-F-2.2).
  • Give each agent only the tools it needs; allowedTools pre-approves, it does not restrict (CCAR-F-2.3).
  • Forced tool_choice returns a 400 on Opus 5.5, Sonnet 5.5 and Fable 5.1; use auto plus a clear instruction or structured outputs.
  • MCP scopes: project .mcp.json for the team, user scope for you (CCAR-F-2.4).
  • Built-in tools: Grep and Glob to find, Read to inspect, Edit over Write for changes (CCAR-F-2.5).
  • Study: 07 · Tool design, 08 · MCP, cheat sheet: MCP.

D3 · Claude Code Configuration & Workflows — 20%

  • CLAUDE.md hierarchy: user, project, subdirectory; keep the root file short (CCAR-F-3.1).
  • Skills (including what used to be custom slash commands) package repeatable team workflows (CCAR-F-3.2).
  • .claude/rules/ with paths frontmatter loads conventions only for matching files (CCAR-F-3.3).
  • Plan mode for unfamiliar or wide changes; direct execution for small, clear ones (CCAR-F-3.4, 3.5).
  • CI: headless claude -p, scoped tools and structured output for review and triage (CCAR-F-3.6).
  • Study: 09 · Claude Code, cheat sheet: Claude Code.

D4 · Prompt Engineering & Structured Output — 20%

  • Explicit criteria cut false positives: say exactly what to flag and what to ignore (CCAR-F-4.1).
  • Few-shot examples for format and edge cases; keep them varied (CCAR-F-4.2).
  • Structured output with JSON schemas and strict tools; validate anyway (CCAR-F-4.3, 4.4).
  • Retry with the validation error fed back; stop after a bounded number of attempts (CCAR-F-4.4).
  • Message Batches for work that can wait; they cannot run your client-side tools mid-request (CCAR-F-4.5).
  • Independent multi-pass review beats one model grading its own work (CCAR-F-4.6).
  • Study: 03 · Prompt engineering, 03 · Structured output, 06 · Batches.

D5 · Context Management & Reliability — 15%

  • Keep critical facts in a stable block outside summaries (CCAR-F-5.1).
  • Escalate on ambiguity or policy limits; don't guess (CCAR-F-5.2).
  • Propagate errors with enough context for the coordinator to decide (CCAR-F-5.3).
  • Large codebases: delegate exploration to subagents, keep notes, compact (CCAR-F-5.4).
  • Human review where confidence is low or impact is high (CCAR-F-5.5).
  • Keep sources and dates attached to claims; report conflicts instead of picking one (CCAR-F-5.6).
  • Study: 04 · Context engineering, 10 · Agent reliability.

Builder's decision reflexes

Situation Reflex
A rule must always hold Enforce it in code or a hook, not only in the prompt
Subagent needs information Put it in the subagent's prompt; it does not see the parent conversation
Agent keeps calling the wrong tool Fix the tool descriptions before adding instructions
Tool fails Return a structured error result so the model can adapt
Output feeds code Schema plus validation plus bounded retry
Long exploration fills context Delegate to a subagent and keep a notes file
Team convention for some files Path-scoped rule in .claude/rules/

Practice: 17 · CCAR-F questions, 18 · Mock exams, 23 · Flashcards.