CCDV-F Exam Preparation Material | Claude Certified Developer - Foundations

Explanation:
Validating each value against business rules catches semantically impossible values that structural JSON parsing cannot see, and quarantining the failures holds them for review before they reach the ledger. Trusting the parsed values ignores that a well-formed structure can still carry invalid data. Retrying re-runs the same fallible extraction without adding any check. Lowering temperature may reduce invalid values but never prevents them and adds no gate.
Reference: https://platform.claude.com/docs/en/build-with-claude/structured-outputs

Explanation:
When a CI/CD rule is non-negotiable, the enforcement point should be outside the model's voluntary behavior. Claude Code hooks are designed as lifecycle automation points that provide deterministic control, so required actions happen, or forbidden actions are blocked, without relying on the LLM to choose correctly every time.
For a rule that must stop an approval before it is posted, PreToolUse is the relevant hook event because it fires before tool execution and can block the call. A command hook can inspect the event-specific JSON it receives on stdin, check whether the passing test artifact exists, and deny the action with feedback when the condition is not met.
PostToolUse is too late for this case because it fires after a successful tool call. A stronger prompt or a UserPromptSubmit hook that appends instructions may improve behavior, but those approaches still rely on model compliance rather than blocking the specific tool call.
See the official Claude Code documentation on Hooks guide and Hooks.

Explanation:
Structured error responses let an
agent make different decisions for different failure modes instead of treating
every backend problem as the same event. In practice, an MCP tool should use an
error signal such asisErrorand include metadata
likeerrorCategory,isRetryable, and a human-readable explanation
that is safe to show or summarize to the customer.
The underlying principle is that recovery logic depends on the nature of the failure: transient errors
may be retried,validation errors usually require corrected
input,business errors such as policy declines should be
explained without retrying, andpermission errors may require
escalation or access handling. Generic failure text prevents Claude from
selecting the right path and often causes wasted retries, premature escalation,
or misleading customer responses.
Fixed retry counts are an anti-pattern when used as the primary response to all failures, because they ignore whether the error is actually retryable. Escalating every tool failure is also too conservative for an 80%+ first-contact resolution target, since many cases can be recovered locally with the right error details.
Learn more about MCP tool behavior in MCP Tools and Claude tool orchestration in Tool Use.

Explanation:
Coordinated subagents pay off on tasks that need wide parallel exploration and more working context than one agent can hold, because separate context windows let workers investigate simultaneously and return compressed findings to an orchestrator. Multi-agent systems consume far more tokens than single-agent sessions, so simple lookups and short summaries do not return that spend. Strictly ordered stages gain nothing from parallel workers. Latency-sensitive streaming replies and fixed-format templated tasks are likewise served better by a single call.
Reference: https://www.anthropic.com/engineering/multi-agent-research-system

Explanation:
Conflicting credible sources require transparent uncertainty handling. When two reliable
sources report different statistics, the synthesis pipeline should preserve
both values, attach source attribution, and explain relevant methodological
differences instead of collapsing the disagreement into a single unsupported claim.
The underlying principle is provenance preservation : downstream report generation should maintain claim-source mappings, methodology context, and conflict annotations so readers can verify and interpret the evidence. This is especially important in multi-agent systems, where summarization can otherwise strip away the reasons that findings disagree.
Selecting the larger-sample statistic, averaging incompatible values, or searching for a convenient tie-breaker all create misleading certainty. These approaches obscure uncertainty, weaken source traceability, and can cause the final report to misrepresent what the evidence actually says.
For implementation patterns involving agent orchestration and structured handoffs between subagents, see the Agent SDK documentation and general guidance in Prompt Engineering.