Question #1

Question 1

A developer is using Claude's structured output to generate JSON configuration files. They need theoutput to include comments explaining each configuration option. JSON doesn't support comments.What is the best structured output approach?

Answer

Answer : D

Why D is Correct? This is the most elegant and practical solution within the constraints of JSON. By designing a schema where each configuration option has both a value and a comment field, you:

Stay within valid JSON — no syntax violations
Keep everything in one response — comments are co-located with their respective config values
Remain machine-readable — the JSON can still be parsed programmatically
Allow downstream processing — a script can strip out comment fields to produce a clean config, while humans can read the annotated version

Example output would look like:

json

{
  "port": {
    "value": 8080,
    "comment": "Server listening port"
  },
  "debug": {
    "value": false,
    "comment": "Enable verbose logging for development"
  }
}

Key Takeaway: When a format has limitations (like JSON lacking comment support), the best approach is to design your schema to work around the limitation — not fight the format or add unnecessary complexity.

Question 2

An AI engineer is implementing an evaluation framework for their agent system. They want to measurewhether the agent asks for clarification when a customer's request is ambiguous rather than makingassumptions. What evaluation metric captures this behavior?

Answer

Answer : B

Why B is Correct? This metric is purpose-built for exactly the behavior being measured. Here's why it's the right fit:

It directly measures the target behavior — whether the agent recognizes ambiguity and responds appropriately by asking for clarification
It uses a ground truth set of known ambiguous inputs, making it objective and repeatable
It produces a quantifiable percentage, making it trackable over time and across model versions
It captures both failure modes:
- Agent asks for clarification when it shouldn't ? false positives
- Agent makes assumptions when it should ask ? false negatives

How it works in practice:

Clarification Rate = (Ambiguous inputs where agent asked for clarification)
                     ?????????????????????????????????????????????????????
                        (Total ambiguous inputs in ground truth set)
                        
                        Target: as close to 100% as possible

Key Takeaway: When evaluating agent behavior, your metric must directly observe that behavior — not infer it from indirect signals like speed or token count. Clarification Rate does exactly that by comparing agent actions against a curated ground truth of ambiguous cases.

Question 3

A developer wants to add a keyboard shortcut in Claude Code that switches between regular mode andplan mode during interactive sessions. What is the default keyboard shortcut for toggling plan mode?

Answer

Answer : A

Explanation:

Shift+Tab is the default keyboard shortcut for toggling plan mode in Claude Code. It cycles through modes — including Plan — mid-session, with no restart required. Lowcode More specifically, Shift+Tab cycles through the three modes in order: Edit ? Auto-Accept ? Plan. Medium Here's why the other options are incorrect:

B. Tab — plain Tab is not a mode-cycling shortcut; it's Shift+Tab.
C. Ctrl+P — this is not a Claude Code shortcut for plan mode. You can type /plan as a slash command, but Ctrl+P is not the keyboard shortcut.
D. Escape — pressing Escape (or Esc twice) is used to cancel or rewind a conversation, not to enter plan mode.

Bonus tip: On Windows with Claude Code 2.1.3+, there is a known bug where Shift+Tab only cycles between Edit and Auto-Accept, skipping Plan mode entirely. The workaround is to use the /plan slash command or the mode selector in the UI instead.

Question 4

A team is building an MCP server that integrates with Salesforce. Their tool catalog includes:search_contacts, get_contact_by_id, update_contact, create_opportunity, get_opportunity_pipeline,update_opportunity_stage, generate_report, and send_email. This is 8 tools. During testing, Claudeoccasionally picks the wrong tool (e.g., search_contacts when get_contact_by_id is more appropriate).What improves tool selection accuracy?

Answer

Answer : A

Explanation:

The root cause of Claude picking the wrong tool is ambiguity in tool descriptions — Claude can't clearly distinguish when to use one over another. Detailed descriptions with explicit usage guidance and negative examples ("NOT for X, use Y instead") directly solve this.

Here's why each option stands:

A ? — Add detailed, differentiating descriptions This is exactly what Anthropic's prompt engineering guidance recommends for tool selection. The example given is ideal because it:

States the specific trigger condition ("when you need to FIND contacts")
Adds a negative example that steers away from the wrong tool ("NOT for retrieving a known contact")
Cross-references the correct alternative ("use get_contact_by_id instead")

This is low-effort, zero architectural change, and directly targets the failure mode observed in testing.

B ? — Reduce to 4 tools by combining operations Merging tools trades one problem for another. A combined search_or_get_contact tool now pushes the disambiguation problem inside the tool call — Claude has to pass a parameter that decides behavior, which is harder to get right and makes tool behavior less predictable. You also lose the clarity of single-responsibility tools.

C ? — Add a tool_router tool This adds latency (an extra round trip on every invocation), burns tokens, and is redundant. The whole point of well-written tool descriptions is that Claude already performs routing natively. Adding a meta-tool to do what descriptions should do is an architectural antipattern that compounds the problem rather than fixing it.

D ? — Rename with numbered prefixes tool1_search, tool2_get conveys zero semantic meaning. Tool names and descriptions are how Claude understands intent — making names less descriptive makes selection worse, not better. This solves nothing.

Question 5

A technical lead is migrating their agent from a 200K token context model (Claude Sonnet 4.5) to a 1Mtoken context model (Claude Sonnet 4.6). They expect to handle longer documents. What otherconsideration should they account for with the larger context?

Answer

Answer : B

Explanation:

The answer is B, but with an important nuance — the specific numbers in the question are partially correct but need clarification. What the official docs actually say: The prompt caching minimum threshold is 1024 tokens for Claude Sonnet 4.5 (and a select group of other models). For Sonnet 4.6 and Opus 4.6, the minimum is 2048 tokens. So option B's core claim — that thresholds differ between the two models — is accurate and real.

Here's a full breakdown:

A ? — "Costs only increase proportionally" This is misleading. Sonnet 4.6 includes the full 1M token context window at standard pricing, so a 900K-token request is billed at the same per-token rate as a 9K request — that part is true. But "only proportionally" ignores the doubled caching threshold (from 1024 ? 2048 tokens), which can silently break an existing caching strategy and cause unexpected cost increases if prompts that were previously cached no longer qualify

B ? — Prompt caching minimums differ This is the correct and most practically important consideration. The minimum token thresholds for prompt caching are: Haiku 4.5 requires 1,024 tokens, while Sonnet 4.6 and Opus 4.6 require 2,048 tokens. A team migrating from Sonnet 4.5 (1024-token minimum) to Sonnet 4.6 (2048-token minimum) could silently lose all prompt caching benefits if their cached content — say, a 1,500-token system prompt — falls below the new threshold. Shorter prompts cannot be cached even if marked with cache_control, and no error is returned, making this a subtle and easy-to-miss regression

C ? — "Just update the model parameter" This is the most dangerous option. Beyond the caching threshold issue, longer context also affects latency, cost per request, and the need to revisit prompt structure. Simply swapping the model string without reviewing caching strategy, cost projections, and context management is a recipe for surprise bills and silent performance regressions.

D ? — "Response quality is identical regardless of context length" False. Research consistently shows that model attention and retrieval accuracy can degrade with very long contexts, particularly for information in the middle of the window (the "lost in the middle" problem). Quality considerations absolutely accompany a 200K ? 1M context migration.

The key takeaway: When migrating to a larger context model, always re-audit your prompt caching strategy. The higher minimum threshold on Sonnet 4.6 means previously-cached prompts may silently stop being cached, turning a cost-saving feature into a hidden cost driver with zero error feedback.

Question 6

A developer has a prompt engineering challenge: they need Claude to fill in a form that has 20 fields,but typically only 5-8 fields are relevant for any given request. The remaining fields should be null.Using structured output with all 20 fields as required causes Claude to hallucinate values for irrelevantfields. What schema design fixes this?

Answer

Answer : B

Explanation:

A is correct. This is a classic structured output problem where the schema design is forcing hallucination.

Here's the full breakdown:

A ? — Make all 20 fields optional (nullable) with clear descriptions The hallucination is caused by required semantics. When a field is marked required, Claude is constrained to produce something — so it invents plausible-sounding values rather than leave a field empty. Removing that constraint resolves it directly. The fix is two-part:

Nullable/optional fields — Claude can now legitimately return null without violating the schema
Clear descriptions indicating when each field applies — this is equally critical. Without guidance like "Only populate if the request involves a physical address", Claude still has to guess. Good descriptions give Claude the decision rule to apply, not just permission to omit

This approach requires zero architectural change, works within a single schema, and keeps the output structure predictable across all requests.

B ? — Dynamic schema generated per-request This sounds sophisticated but introduces real problems. Generating a schema per-request means you need a first pass to analyze the input and decide which fields are relevant — which is exactly the reasoning you want Claude itself to do during form-filling. You've now added latency (an extra LLM call or classification step), complexity in schema management, and a new failure mode: what if the schema generator gets it wrong? You've also lost a single predictable output shape, complicating downstream parsing. Option A achieves the same selective population without any of this overhead.

C ? — Add a confidence field alongside each This doubles your schema complexity (20 fields ? 40 fields) without solving the root cause. Claude still has to produce a value for required fields — now it just also rates how confident it is in the hallucination. Downstream, you'd need logic to discard low-confidence values, which means you're building the null-handling you could have gotten for free with option A. It also makes the output significantly harder to parse and use.

D ? — Create 4 schemas for different form types This only works if your form types are perfectly discrete and known in advance — which the question implies they aren't (5–8 variable fields relevant per request). You'd also need a classification step to route to the right schema, and edge cases that span multiple "types" would fall through the cracks. It trades one maintenance problem (hallucinated fields) for another (schema proliferation and routing logic).

Option A handles the variability naturally. The core principle: Schema constraints are instructions. Marking a field required tells Claude it must have a value, which guarantees hallucination when the data doesn't exist. Making fields optional with clear applicability descriptions tells Claude when to populate them — which is the actual behavior you want.

Question 7

A solutions engineer is troubleshooting an agent that uses the Agent SDK. The agent should use the'search' tool before answering questions, but it frequently answers from training knowledge withoutsearching. The tool is properly defined and the system prompt says to 'use the search tool to findanswers.' What is the most effective fix?

Answer

Answer : A

Explanation:

A ? — Force tool_choice: {type: 'tool', name: 'search'} on the first turn, then switch to auto This precisely matches the failure mode. The agent is skipping the search on the initial turn — answering directly from training knowledge before any tool use occurs. Forcing tool_choice to a specific tool on turn 1 makes that search non-negotiable at the API level, not just at the instruction level. Then switching to auto for subsequent turns is the right design because:

Follow-up reasoning, synthesis, and clarifying questions shouldn't require a search every time
The model needs freedom to call other tools or respond directly once it has retrieved real information
Locking tool_choice permanently would cause unnecessary searches on turns where they add no value

This is a surgical, architectural fix — it solves the problem exactly where it occurs (turn 1) without over-constraining the rest of the conversation.

The core principle: When instruction-following fails for a specific behavioral requirement, move enforcement from the prompt layer to the API constraint layer. tool_choice with a named tool is a hard constraint — the model cannot route around it regardless of what its training knowledge suggests. Reserve that constraint for exactly the turn where the behavior is required, then release it.

Question 8

A team maintains a Claude Code project with a growing CLAUDE.md file that's now 800 lines. ClaudeCode seems to be ignoring rules defined in the latter half of the file. What is the recommendedapproach for managing large CLAUDE.md files?

Answer

Answer : D

Explanation:

D is correct — and confirmed by official Claude Code documentation and best practices.

D ? — Concise root CLAUDE.md + @import for detail + .claude/rules/ with glob patterns

This directly addresses the root cause. If your CLAUDE.md is too long, Claude ignores half of it because important rules get lost in the noise. The fix operates on two levels:

Structural decomposition: The .claude/rules/ directory lets you organize project instructions into multiple focused rule files instead of one large CLAUDE.md. All markdown files in this directory are automatically loaded into Claude Code's context when launched.

Context-aware loading via glob patterns: The most powerful aspect of .claude/rules/ is conditional application — you can scope rules to specific files using YAML frontmatter with a paths field. This means React component rules only load when working on React files, database rules only when touching database files, and so on — the right rules are injected at the right moment rather than everything competing for attention all the time

Priority is preserved: Rules files load with the same high priority as CLAUDE.md. When a rule has paths frontmatter, it only loads (and receives high priority) when Claude is working on matching files

The practical structure looks like:

your-project/
??? CLAUDE.md              # Critical rules only, ~50-100 lines
??? .claude/
    ??? rules/
        ??? testing.md     # Auto-loaded everywhere
        ??? security.md    # Auto-loaded everywhere
        ??? frontend/
            ??? react.md   # paths: src/**/*.tsx — loads only for React files

Question 9

A developer builds an MCP tool that generates images using Stable Diffusion. The tool takes a promptand returns a base64-encoded image. Claude then describes the image to the user. However, Claude'sdescriptions don't match the generated images because Claude can't actually see the tool output as animage. How should this be architectured?

Answer

Answer : D

Explanation:

D is correct. This is fundamentally a data type problem — Claude is receiving image data encoded as a string, not as an image content block, so it literally cannot see it. D ? — Include the base64 image in a subsequent message with the image content type The Anthropic API supports multimodal content blocks. A tool result can include — or a follow-up message can contain — an image content block with source.type: "base64" and the correct media_type. When structured this way, Claude actually processes the image through its vision capabilities rather than seeing a wall of base64 text. The correct architecture:

Tool runs, generates the image, returns base64
The MCP layer or orchestrator wraps that base64 in a proper image content block
Claude receives it as an actual image and can genuinely see and describe it

This is the only option that gives Claude real visual grounding — everything else is a workaround that produces descriptions disconnected from actual output.

The core principle: Claude's multimodal capabilities only activate when image data is passed through the correct content type structure in the API. Base64 as a raw string is opaque text. Base64 wrapped in {"type": "image", "source": {"type": "base64", "media_type": "image/png", "data": "..."}} is a real image Claude can see. The architecture fix is in how the data is typed and delivered, not in working around the limitation.

Question 10

Your team is designing an agent evaluation framework. They need to test whether the agent handlestool failures gracefully across 50 different error scenarios. Running each scenario manually takes 10minutes. What is the most efficient approach?

Answer

Answer : D

Explanation:

D — Message Batches API with mocked tool failures, analyzed programmatically This is purpose-built for exactly this use case. The math alone makes the case: 50 scenarios × 10 minutes each = 8+ hours of sequential manual testing. The Batches API runs all 50 in parallel as a single async job at a 50% token cost discount, and returns structured results you can analyze programmatically — pass/fail rates, failure mode categorization, response pattern analysis — without any human in the loop. The architecture is clean:

Each of the 50 scenarios is a self-contained request with a system prompt defining the agent, a user message triggering the tool call, and a mocked tool result returning the specific error condition to test
The batch job runs them all concurrently overnight or in the background
Results come back as structured output you diff against expected graceful-handling behavior
Regressions are caught automatically on every future code change by re-running the batch

This is how evaluation frameworks are supposed to work — deterministic, repeatable, scalable, and cheap enough to run continuously.

Anthropic CCA-F Exam Questions Answers

PDF Only

$59

Test Engine

$69

PDF + Test Engine

$79

Anthropic CCA-F Last 24 Hours Result

Anthropic CCA-F Practice Test Questions ( Updated) – Real Exam Questions & Dumps PDF

Latest Anthropic CCA-F Dumps PDF (Updated )

Verified CCA-F Exam Questions and Answers

Real Exam Simulation with Practice Test Engine

Why Certs4sure Is the Right Choice for CCA-F Exam Preparation

Question #1

Question #2

Question #3

Question #4

Question #5

Question #6

Question #7

Question #8

Question #9

Question #10

Leave Your Review