Certification Study Guide

Claude Certified Architect
Foundations (CCA-F)

Everything you need to pass the CCA-F exam: domain breakdowns, key patterns, anti-patterns, and interactive practice questions.

5

Exam Domains

720

Passing Score

30

Practice Questions

Exam Blueprint

The CCA-F is a proctored, multiple-choice exam with a scaled passing score of 720/1000. You'll face 4 of 6 possible scenarios.

67% rule: D1 + D3 + D4 account for two-thirds of your score. Prioritize accordingly.

The 6 Scenarios

You'll face 4 of these, randomly selected. Know all 6.

1. Customer Support Agent

Escalation, refund tools, 80%+ FCR target

D1D2D5

2. Code Gen with Claude Code

CLAUDE.md, slash commands, plan mode

D3D5

3. Multi-Agent Research

Coordinator + subagents architecture

D1D2D5

4. Developer Productivity

Codebase exploration, built-in tools, MCP

D2D3D1

5. Claude Code for CI/CD

Automated code review, test gen, PR feedback

D3D4

6. Structured Data Extraction

JSON schema extraction, validation, edge cases

D4D5

D1 · 27%

Agentic Architecture & Orchestration

The highest-weighted domain. Master the agentic loop, coordinator-subagent pattern, and the deterministic vs. probabilistic distinction.

The Agentic Loop

Anti-Patterns

Parsing natural language to detect completion
Arbitrary iteration caps as primary stop
Checking assistant text for "I'm done"

Correct Pattern

Always use stop_reason to control the loop
"tool_use" = execute and continue
"end_turn" = stop and return

The Core Exam Distinction

This pattern appears across almost every domain. If the question involves money, identity, or security — the answer is programmatic enforcement.

Approach	Guarantee	Use When
Programmatic enforcement (hooks, prerequisite gates)	Deterministic — 100%	Critical business logic, financial ops, identity verification
Prompt-based guidance (system prompt, few-shot)	Probabilistic — non-zero failure	Judgment calls, tone, escalation phrasing

Hub-and-spoke: The coordinator manages ALL inter-subagent communication. Subagents have isolated context — they don't inherit the coordinator's conversation history.

Coverage gaps? Root cause is the coordinator's task decomposition being too narrow, not the subagents failing. Fix: iterative refinement (evaluate → find gaps → re-delegate).

Parallel execution: Emit multiple Task tool calls in a single coordinator response — not separate turns.

Hooks: PostToolUse hooks transform tool results (date normalization). Tool call interception hooks block outgoing calls (refund > $500). Know the difference.

D2 · 18% D3 · 20%

Tool Design & Claude Code Config

Tool descriptions are the #1 mechanism for LLM tool selection. CLAUDE.md hierarchy controls what Claude knows about your project.

Tool Description Quality

Bad Description

Good Description

get_customer: "Retrieves customer information"
lookup_order: "Retrieves order details"

Result: Agent frequently calls get_customer when users ask about orders.

get_customer:
  "Retrieves customer profile by ID or email.
   Returns: name, email, account status.
   Use when: user asks about account, profile, identity.
   Do NOT use for: order lookups, billing history."

lookup_order:
  "Retrieves order details by order number.
   Returns: items, status, tracking, dates.
   Use when: user asks about orders, delivery, returns.
   Do NOT use for: customer profile info."

Clear boundaries + negative examples eliminate misrouting.

`tool_choice` Options

Setting	Behavior	Use When
`"auto"`	Model may skip tool call entirely	Default — model decides
`"any"`	Must call a tool (chooses which)	Need structured output, not text
`{"type":"tool","name":"X"}`	Must call specific tool X	Enforce a first step (e.g., extract_metadata)

CLAUDE.md Hierarchy

Exam trap: new team member doesn't get instructions because they're in ~/.claude/CLAUDE.md (user-level) instead of .claude/CLAUDE.md (project-level).

Level	Location	Shared?	Scope
User	`~/.claude/CLAUDE.md`	No	All projects for that user
Project	`.claude/CLAUDE.md`	Yes (VC)	All team members
Directory	Subdirectory `CLAUDE.md`	Yes	Specific subdirectory
Path rules	`.claude/rules/*.md`	Yes	Files matching glob pattern

Grep searches file contents. Glob matches file names. Read loads full files. Edit makes targeted changes via unique text match. When Edit fails (non-unique match), fall back to Read + Write.

MCP scoping: Shared team servers go in .mcp.json (version controlled). Personal/experimental in ~/.claude.json. Use ${GITHUB_TOKEN} syntax for credentials.

Skills vs. CLAUDE.md: Skills are on-demand (context: fork isolates verbose output). CLAUDE.md is always-loaded. Path-specific rules (.claude/rules/) load only when editing matching files — more efficient than per-directory CLAUDE.md.

D4 · 20%

Prompt Engineering & Structured Output

Vague instructions fail. Specific categorical criteria work. tool_use eliminates JSON syntax errors but not semantic errors.

Explicit Criteria for Precision

Vague (Fails)

"Be conservative"
"Only report high-confidence findings"
"Check that comments are accurate"

Specific (Works)

"Flag only: bugs, security issues. Skip: style, local patterns"
"Report when claimed behavior contradicts actual code"
"Flag comments only when they contradict the code"

Structured Output Reliability

tool_use Eliminates

JSON syntax errors
Malformed output
Missing required fields

tool_use Does NOT Fix

Line items that don't sum to total
Fabricated values for required fields
Data in wrong semantic fields

Schema tip: Make fields optional/nullable when source documents may not contain the info. Required fields cause the model to fabricate values.

Batch API Facts

50%

Cost Savings

24h

Max Processing

No

Multi-turn Tools

custom_id

Correlation

Good for overnight reports. Bad for blocking pre-merge checks.

D5 · 15%

Context Management & Reliability

Keep critical information alive across long conversations, handle errors gracefully, and know when to escalate.

Three Context Killers

1

Progressive Summarization

Condensing "$127.43 refund for order #8821" into "discussed refund" loses critical details.

2

Lost in the Middle

Models reliably process beginning and end but may omit findings from the middle.

3

Tool Result Bloat

40+ fields per tool result when only 5 are relevant. Trim before accumulation.

Fix: Extract transactional facts into a persistent "case facts" blockA structured block of key facts included in every prompt, separate from summarized conversation history in each prompt.

Escalation Rules

Escalate When

Customer explicitly requests a human
Policy exceptions or gaps
Unable to make meaningful progress

Unreliable Signals

Sentiment-based escalation
Self-reported confidence scores
Frustration level alone

Exam rule: Customer says "I want to talk to a real person" → honor immediately. Even if the issue is simple.

Error Categories

Category	Retryable?	Example
Transient	Yes	Timeout, service unavailable
Validation	No	Invalid input format
Business	No	Policy violation (refund > limit)
Permission	No	Unauthorized access

Test Your Knowledge

5 exam-style questions spanning all domains.

D1 · Question 1 of 5

Your customer support agent occasionally processes refunds without first verifying customer identity. The error rate is 8%. What's the most effective fix?

When a specific tool sequence is required for critical business logic (financial operations), programmatic enforcement provides deterministic guarantees. Prompt-based approaches have non-zero failure rates.

Study Complete

You've Got the Blueprint

You've covered all 5 domains, the core exam patterns, and tested yourself with practice questions.

Priority

Score 900+ on Practice Exam

Focus

D1 is 27% of the exam

Pattern

Deterministic > Probabilistic

Key

Know anti-patterns cold

Claude Certified ArchitectFoundations (CCA-F)

Exam Blueprint

The 6 Scenarios

1. Customer Support Agent

2. Code Gen with Claude Code

3. Multi-Agent Research

4. Developer Productivity

5. Claude Code for CI/CD

6. Structured Data Extraction

Agentic Architecture & Orchestration

The Agentic Loop

Anti-Patterns

Correct Pattern

The Core Exam Distinction

Tool Design & Claude Code Config

Tool Description Quality

tool_choice Options

CLAUDE.md Hierarchy

Prompt Engineering & Structured Output

Explicit Criteria for Precision

Vague (Fails)

Specific (Works)

Structured Output Reliability

tool_use Eliminates

tool_use Does NOT Fix

Batch API Facts

Context Management & Reliability

Three Context Killers

Progressive Summarization

Lost in the Middle

Tool Result Bloat

Escalation Rules

Escalate When

Unreliable Signals

Error Categories

Test Your Knowledge

D1 · Question 1 of 5

You've Got the Blueprint

Claude Certified Architect
Foundations (CCA-F)

`tool_choice` Options