Claude Certified Architect – Foundations

1-Day Before Exam Review Notes

Use this as a final refresher, not as a substitute for the official exam guide, Anthropic training, or hands-on practice. Focus on distinctions and decision patterns rather than memorizing wording.

Aligned to the current CCAR-F exam blueprint · Updated September 2026

Start Free Mock Exam Official Certification Page ↗

First: remember the exam shape

27%

Agentic Architecture & Orchestration

18%

Tool Design & MCP Integration

20%

Claude Code Configuration & Workflows

20%

Prompt Engineering & Structured Output

15%

Context Management & Reliability

1. Agentic Architecture & Orchestration

27%

Choose architecture based on task shape: fixed workflow for predictable, repeatable steps; autonomous agents when the path depends on intermediate findings; multi-agent systems when distinct specialties, tools, or parallel work justify the coordination cost.
The agentic loop lifecycle: send request → inspect stop_reason → if tool_use, execute tool calls and return tool_result content blocks → repeat; if end_turn, exit loop. Anti-patterns: parsing natural-language signals to determine termination, setting arbitrary iteration caps as the primary stopping mechanism, or checking for assistant text content as a completion indicator. stop_reason is the only reliable signal.
stop_reason values: end_turn = model finished naturally; tool_use = model paused, tool calls must be executed and results returned; max_tokens = response truncated and incomplete, never forward as complete; stop_sequence = custom stop string matched.
Parallelism: a coordinator emits multiple Task tool calls in a single response to spawn parallel subagents. Emitting them across separate turns means sequential execution. Parallelism comes from the single-response batch, not from the number of turns.
Session state: --resume with a session name continues a prior session when the prior context is mostly still valid. fork_session creates a parallel exploration branch from the current state. Choose fresh session with injected summaries when prior tool results are stale or the context has drifted.
Hooks give deterministic guarantees that prompt instructions cannot: PostToolUse hooks normalize heterogeneous data formats after every tool call regardless of model behavior; interception hooks block policy-violating calls outright (e.g. refunds above a threshold). Hooks vs prompts = deterministic enforcement vs probabilistic compliance.
Delegation must be operational: the coordinator needs access to the Task capability, subagents need clear descriptions, focused prompts, and appropriate tool restrictions. Subagents are isolated — do not assume conversation history or memory carries across invocations; pass required context explicitly.
On escalation to a human who cannot see the transcript, emit a structured handoff summary: what was attempted, what succeeded, what failed, open questions, and any constraints or deadlines. An unstructured narrative forces the human to reconstruct context.
Use iterative refinement when quality must improve through evaluate → targeted follow-up → re-synthesize loops instead of accepting the first pass.

2. Tool Design & MCP Integration

18%

Match the tool to the job: Glob finds files by path pattern (e.g. **/*.tsx) — use it when you know the name or extension but not the location. Grep searches file contents by regex — use it when you know a symbol, string, or concept exists somewhere in the codebase. Never use Glob to find content or Grep to find file names.
Prefer clear, non-overlapping tool descriptions and smaller role-relevant toolsets. More tools are not automatically better if they increase selection ambiguity. Keyword-sensitive instructions in the system prompt can create unintended tool associations — keep tool descriptions tightly scoped and rename or split tools when overlap causes mis-selection.
tool_choice controls which tools the model may call: 'auto' lets the model decide; 'any' forces a tool call but allows the model to pick which one; forced tool selection names a specific tool the model must call. Use forced selection when the workflow requires a guaranteed tool call at a specific step.
MCP structured errors: use the isError flag pattern to signal tool failures as data rather than exceptions. Error categories — transient (retry), validation (fix input), business (policy block), permission (authorization failure) — and an isRetryable boolean let the orchestrator choose the right recovery path without parsing error text.
Use MCP resources for discoverable catalogs and reference content rather than forcing an agent to discover everything through exploratory tool calls. Prefer existing community MCP servers over custom implementations for standard integrations (e.g. Jira, GitHub, Slack) — custom servers add maintenance burden without providing differentiation.
Keep shared MCP configuration at project scope and credentials out of source control; personal or experimental servers should remain personal rather than being imposed on the team.

3. Claude Code Configuration & Workflows

20%

Memory hierarchy (outermost to innermost, all active simultaneously): user-level ~/.claude/CLAUDE.md for personal cross-project preferences; project-level CLAUDE.md for team-wide conventions; directory-level CLAUDE.md files for path-specific rules. Use @import to pull in additional files. Use the /memory command to verify exactly which memory files are currently loaded.
Project-wide, always-relevant conventions belong in project CLAUDE.md. Specialized instructions that apply only to certain paths belong in .claude/rules/ with path globs — importing a file into CLAUDE.md does not make it conditionally path-scoped.
Commands live in .claude/commands/ (project-scoped) or ~/.claude/commands/ (user-scoped). Skill frontmatter controls behavior: context: fork isolates the skill's context from the main session; allowed-tools restricts which tools the skill may use; argument-hint documents what argument the skill expects. Use skills for on-demand, task-specific workflows rather than loading occasional procedures into every session.
CI/CD non-interactive mode: use -p (or --print) to run Claude Code without interactive prompts — suitable for pipelines. --output-format json produces machine-readable output; --json-schema passes a schema for structured output. Critical: use a separate, fresh Claude instance for code review — the same session that generated the code is less effective at catching its own errors.
The Explore subagent isolates verbose file-discovery output from the main session context. Delegate broad codebase searches to it so large result sets do not consume the main context window.
Iterative refinement patterns: provide concrete input/output examples rather than abstract descriptions; iterate test-first so failures are immediately visible; use the interview pattern — have Claude ask clarifying questions to surface considerations the developer may not have anticipated; batch interacting fixes (changes that affect each other) but sequence independent fixes to keep diffs reviewable.
Plan mode is valuable for large or architectural changes; direct execution is more appropriate for small, well-scoped edits.
Read, Write, and Edit have distinct contracts. Read retrieves file contents — use offset and limit for large files to avoid wasting context. Write creates a new file or completely overwrites an existing one — use only for new files or full rewrites. Edit makes a targeted old-string → new-string replacement in an existing file — prefer it for any modification. When Edit fails due to non-unique text matches, fall back to Read + Write. Bash is the last resort for file work; use the dedicated tools first.

4. Prompt Engineering & Structured Output

20%

Few-shot prompting: provide 2–4 targeted examples for ambiguous scenarios. Examples must show reasoning for why one action was chosen over plausible alternatives — not just the correct answer. Examples that only show inputs and outputs without rationale do not transfer judgment.
Schema mechanics: tool_use with JSON schema is the reliable path to structured output. Mark fields optional (nullable) when source documents may not contain the information — never let the model fabricate values to satisfy a required field. Use enum patterns with an 'unclear' or 'other' + detail field to handle cases outside the expected value set rather than forcing a best guess.
Structured output guarantees schema conformance, not semantic correctness, cross-record consistency, or truth against an external authoritative source. Use deterministic validation for constraints the model should not be trusted to self-certify. Retries are effective for transient model errors but ineffective when the required information is simply absent from the source document — detect absence explicitly. Self-check fields (e.g. calculated_total vs stated_total) catch arithmetic and consistency errors without re-running the full extraction.
Batch API: use when latency tolerance allows — up to 50% cost saving, processing window up to 24 hours, no guaranteed latency SLA. Supply a custom_id per request for correlation. The Batch API does not support multi-turn tool calling within a single request. Use it for bulk extraction, classification, or evaluation jobs; use the synchronous API when the caller needs a response before proceeding.
Multi-instance review: separate the generator instance from the reviewer instance — the model that produced the code is less effective at reviewing its own output. Structure the review as per-file passes followed by a cross-file pass to catch inter-file inconsistencies. Preserve independent first-pass judgments before instances see each other's output; adjudicate real disagreements against an explicit rubric.
Treat prompt and workflow structure as a control surface: clear role boundaries, explicit success criteria, and examples are often more reliable than vague instructions to 'be careful.'

5. Context Management & Reliability

15%

Escalation triggers: a customer explicitly requesting a human agent; a policy exception or gap the system cannot resolve; inability to make meaningful progress after reasonable attempts. Unreliable escalation proxies: negative sentiment and self-reported confidence scores do not reliably indicate actual case complexity. When multiple customer records match, ask for an identifier — never guess which record to act on.
Context hygiene: maintain a persistent 'case facts' block at the top of context for amounts, dates, IDs, and key decisions so they survive compaction. Trim verbose tool outputs before passing them forward — summarize, don't relay. The 'lost in the middle' effect means models attend less to content in the center of a long context; put key summaries first or last.
When stop_reason is max_tokens the model was cut off mid-response. Never forward a truncated output downstream as if it is complete — increase the max_tokens budget, compact or summarise context to free space, or surface a degraded/partial result.
When context becomes noisy during a long session, /compact can summarize the conversation so work can continue with less context pressure. Keep durable findings in a scratchpad or structured artifact and delegate verbose investigation to subagents that return concise summaries.
Human review accuracy: aggregate accuracy metrics (e.g. 97% overall) may mask poor performance on specific document types or fields. Use stratified sampling of high-confidence outputs to catch systematic errors that overall metrics hide. Calibrate field-level confidence thresholds on labeled sets rather than relying on global model confidence.
Do not confuse a failed tool or unavailable provider with a valid empty result. Preserve the distinction so downstream decisions remain honest. After bounded retries are exhausted, preserve successful partial results, mark the overall outcome as degraded or partial, and block claims that depend on missing evidence.
Provenance must survive every handoff: carry stable claim/source identifiers through summarization. When sources conflict, annotate with attribution rather than picking one silently. Require publication dates for time-sensitive claims. Maintain crash-recovery manifests so a resumed pipeline knows which steps completed.

High-value distinctions to remember

stop_reason-driven vs text-driven loop termination: inspect stop_reason after every response — it is the only reliable signal. Never parse assistant text content or natural-language phrases to decide whether the loop is done.
Hooks vs prompts: hooks enforce policy deterministically (PostToolUse normalisation, interception blocks); prompt instructions produce probabilistic compliance. Use hooks when the guarantee must hold regardless of model behavior.
resume vs fresh-with-summary: --resume continues a prior session when prior context is mostly valid; start fresh with an injected summary when prior tool results are stale or context has drifted significantly.
Sync API vs Batch API: synchronous API when the caller must wait for a response; Batch API for latency-tolerant bulk jobs (up to 50% cheaper, up to 24-hour window, no latency SLA, no multi-turn tool calling per request).
Escalation trigger vs sentiment/confidence proxy: escalate on explicit customer request, policy gap, or inability to progress — not on negative sentiment or self-reported low confidence, which do not reliably correlate with case complexity.
stop_reason end_turn vs tool_use vs max_tokens: end_turn = done; tool_use = execute tools and continue; max_tokens = truncated, incomplete; stop_sequence = custom stop string hit.
Fixed workflow vs agent: known sequence vs path that emerges from evidence.
Single agent vs multi-agent: one focused context vs distinct specialties/tools/context or useful parallelism.
CLAUDE.md vs rules vs skills: always-active project guidance vs path-scoped automatic guidance vs on-demand workflow.
Write vs Edit: Write creates or fully overwrites — all existing content is replaced. Edit patches a specific old-string → new-string — everything else is preserved. When Edit fails due to non-unique matches, fall back to Read + Write.
Glob vs Grep: Glob matches file paths/names by pattern; Grep matches file contents by regex. Using the wrong one means missing the result entirely.
Schema-valid vs correct: structured output can be well-formed and still be wrong or fabricated.
Empty result vs failed result: 'nothing found' is not the same as 'search unavailable.'
Summary vs provenance: compression is fine only if claim-to-source lineage survives; conflicting sources must be attributed, not silently resolved.

Final 15-minute checklist

Review the five domain names and weights so you know where the exam emphasis sits.
Revisit mistakes from your most recent mock exam instead of starting new material.
Mentally explain the fourteen distinctions on this page without looking at the answers — especially stop_reason loop termination, hooks vs prompts, Batch API, and escalation triggers.
Confirm exam logistics, identification, time, delivery method, and current policies on the official page.
Stop studying early enough to rest. The goal now is recall and judgment, not adding more topics.

About these notes: They summarize recurring concepts and decision patterns from the AI Pathway CCAR-F preparation sources and current exam blueprint. Always use the official Anthropic certification page for the latest exam policies, registration details, and blueprint changes.