Claude Certified Architect — Foundations Cheat Sheet (CCAR-F)

Read the ask before the options: note how many answers it wants ("Choose TWO") and whether it's phrased negatively ("Which will NOT…").

Item counts are 60 × weight, rounded. The guide gives weights only. Real forms vary.

Spot which scenario you're in from its first item and you can predict the topics:

Most items come down to one of these:

• Money, security, compliance, "100%" → enforce in code (hook, prerequisite gate, schema). Prompts and few-shot get you "usually right", never guaranteed.
• Fix the root cause the evidence shows → it is often upstream (coordinator, input). Don't change a component that works.
• Proportionate first step → cheapest fix first: tool description → explicit criteria + few-shot → routing layer → classifier.
• Least privilege → 4–5 tools per agent, by role. Add a narrow tool for a frequent cross-role need.
• Structured beats generic → errors carry a category and isRetryable; handoffs have fixed fields; output follows a JSON schema.
• Isolate context → subagents inherit nothing, so pass what they need. Reviewers are separate instances.
• Don't trust self-reported signals → self-rated confidence, sentiment and "I'm done" text are unreliable. Use stop_reason and explicit criteria.
• A bigger context window doesn't fix attention dilution → split into passes, put key facts first, keep a scratchpad.
• Never hide failures, never kill the whole workflow → recover locally, then propagate with partial results. A valid empty result is not an error.
• Preserve information, don't let the model invent it → nullable fields, case facts block, claim → source, keep both values when sources conflict.
• Ask instead of guessing → several matching records: ask for another identifier. Policy is silent: escalate.
• Declarative and automatic beats manual → path rules over a giant CLAUDE.md; project config over reminding each person.
• Only documented features → -p, .mcp.json, .claude/rules/. Not CLAUDE_HEADLESS, --batch or .claude/config.json.

D1 · Agentic Architecture & Orchestration (27%)

flowchart LR
  A[Call Messages API] --> B{stop_reason}
  B -- tool_use --> C[Run tool]
  C --> D[Append tool_result to history]
  D --> A
  B -- end_turn --> E[Done]
  B -- max_tokens --> F[Truncated, not finished]

• Loop control → branch on stop_reason, not on content[0].type or whether text exists. An iteration cap is only a safety net.
• Claude "can't see" the tool result → the tool_result was never appended to the conversation.
• Report misses a whole topic category → the coordinator decomposed too narrowly. Fix decomposition, not the subagent.
• Subagents talk to each other → wrong. Hub-and-spoke: everything goes through the coordinator.
• Coordinator can't spawn subagents → allowedTools lacks "Task" ⚠ (newer docs call it the Agent tool).
• Synthesis loses citations → metadata was stripped at the coordinator boundary. Pass structured findings with source and date.
• Refunds sometimes skip verification → prerequisite gate: process_refund runs only once a verified customer ID exists.
• Human handoff → self-contained: customer ID, summary, root cause, amount, recommendation. The human has no transcript.

• 3 of 50 files changed → tell the resumed session which files changed and re-analyse only those.
• Fixed pipeline vs dynamic → known steps, structured input: prompt chaining. Open-ended investigation: dynamic decomposition.

D2 · Tool Design & MCP Integration (18%)

• Model confuses two tools → expand descriptions first: purpose, input format, examples, edge cases, "use X instead of Y when…".
• Good descriptions, still misrouted → look for keyword rules in the system prompt that override them.
• One tool does too much, or names overlap → split into narrow tools with clear names.
• 18+ tools on one agent → cut to 4–5 by role.

• Structured error → isError: true + errorCategory + isRetryable + a human-readable description. Generic "Operation failed" is always wrong.
• No results → isError: false with 0 results is a valid answer, don't retry. An unreachable source is an error, never an empty list.

• Jira, GitHub, other standard integrations → community MCP server first; build custom only for unique needs.
• Grep vs Glob → Grep searches file contents (callers, imports, error strings); Glob matches paths and names.
• Edit fails on a non-unique anchor → fall back to Read + Write ⚠.

D3 · Claude Code Configuration & Workflows (20%)

Levels are combined, not overridden. Check what loaded with /memory ⚠ (newer versions also show it in /context).

• Skill frontmatter → context: fork runs it in an isolated subagent (good for verbose output), allowed-tools restricts tools, argument-hint prompts for parameters.
• Plan mode vs direct → decide by ambiguity, not size: architecture, many files, several valid approaches → plan; clear bug with a stack trace → direct.
• Refinement → model misreads the requirement: 2–3 input/output examples. Tricky edge cases: write tests first. Unfamiliar domain: interview pattern.
• Several fixes → interdependent: one message. Independent: one at a time.

• Reviews repeat old comments → put prior findings in the prompt; report only new or unresolved issues.
• Reviewing code Claude just wrote → a separate instance, never the generating session.

D4 · Prompt Engineering & Structured Output (20%)

• Too many false positives → concrete report/skip criteria + code examples per severity. "Be conservative" is wrong.
• One category at ~40% false positives drags trust in all findings → temporarily disable that category, fix its criteria, re-enable.
• Few-shot done right → 2–4 examples targeting ambiguous cases, with the reasoning, varied document structures ⚠.
• Few-shot can't fix → malformed JSON (use schema), invented values (nullable), wrong tool (description), wrong totals (validation).
• tool_use + JSON schema → guarantees structure, not correctness. Strict mode removes syntax errors.
• Model invents absent values → make the field optional/nullable; add "other" + detail and "unclear" to enums.
• Validation-retry → resend the document + the failed output + the specific error. Pointless if the data isn't in the source; return null or route to a human.
• Semantic checks → extract calculated_total alongside stated_total; flag conflict_detected.

• Batch partly fails → resubmit only the failed custom_ids, with a change (e.g. chunk oversized documents).
• 30h SLA with a 24h batch window → 6h of slack, so submit about every 4 hours.
• Confidence scores → use them for routing, calibrate them; they never replace criteria.

D5 · Context Management & Reliability (15%)

• Amounts, dates, IDs lost after summarising → a structured case facts block included in every prompt, never summarised.
• Lost in the middle → key findings first, section headers; reorder the input rather than repeat it.
• Tool returns 40 fields → trim to the needed ones before they enter history.
• Escalation triggers → customer asks for a human (immediately) · policy is silent or an exception applies · no progress after a genuine attempt.
• Not triggers → frustration on a simple case, low self-rated confidence. Resolve it.
• Several customers match → ask for email, phone or order number; never pick the most recent.
• FCR 55% vs an 80% target → explicit escalation criteria + few-shot before any classifier.
• Error propagation → report failure type, what was tried, partial results, alternatives. A timeout returned as results: [] with success is silent suppression.
• Answers drift into generic "typical patterns" → context degradation: scratchpad file, subagents for narrow questions, /compact early.
• Long exploration may crash → state manifest: phase, explored paths, findings, next steps.
• 97% overall accuracy → can hide a weak segment. Measure by document type and field; stratified sampling includes high-confidence output.
• Two credible sources conflict → keep both values with sources. Don't pick the newest or average them; different dates may be a trend.
• Every report claim → claim + source + excerpt + date, plus a coverage gap section for sources that couldn't be checked.

Wrong options tend to follow a few patterns:

Out of scope per the guide. An option built on these is usually a distractor: fine-tuning or training, API auth/billing/accounts, OAuth and key rotation, deploying or hosting MCP servers, Claude's internals or training (RLHF, Constitutional AI), embeddings and vector DBs, computer use, vision, streaming/SSE, rate limits and pricing math, cloud-provider configs, benchmarks, tokenization, and prompt caching beyond knowing it exists.

• Hard constraints (money, security, compliance) go in code; everything else starts with the cheapest fix that targets the root cause.
• stop_reason drives the loop; tool_result must be appended; subagents get context only through their prompt.
• Structure everywhere: typed errors with isRetryable, fixed handoff fields, JSON schema with nullable fields, a case facts block.
• Scope matters: project (.claude/, .mcp.json) for teams, user (~/.claude/) for personal setup.
• Batch for anything nobody waits on; synchronous for anything that blocks a person.