Claude Certified Developer Foundations (CCDV-F) Practice & Architecture

Comprehensive vocabulary flashcards covering architectural patterns, prompt engineering, agent systems, Claude Code configurations, and security protocols for the Claude Certified Developer Foundations (CCDV-F) certification.

Claude Agent SDK

An official SDK that abstracts the agent execution loop, tool dispatch, and conversation history management into pre-built interfaces rather than requiring developers to write a custom harness from scratch.

Skills Directory

A designated directory recognized by Claude Code where custom skills consisting of instructions and support scripts are defined so Claude can discover and load them dynamically during repository sessions.

Format Drift

A failure mode occurring over multi-turn conversations where model outputs gradually lose their expected structured schema (such as JSON); remediated through context management, prompt adjustments, or output validation.

Defensive Parsing

An output consumption technique designed to gracefully handle unexpected fields, missing values, and type mismatches in model responses so downstream systems do not crash.

Session Hygiene

The practice of starting a fresh session or explicitly clearing context between independent, unrelated tasks to prevent historical conversation details from contaminating new outputs.

Fixed Workflow Pattern

An architecture where work follows a predefined, identical sequence of discrete LLM calls without runtime autonomous planning, offering high predictability, simple debugging, and low cost.

Supervisor Agent

A top-level agent in a hierarchical multi-agent architecture responsible for decomposing user requests, dispatching tasks to specialized subagents, and aggregating or reconciling their individual outputs.

Deterministic Hook

Code that executes unconditionally on specified system events (such as file edits or command execution) outside the model's discretion, guaranteeing policy enforcement that probabilistic prompts cannot assure.

Loop Bounding ($\text{max\_turns}$ and $\text{max\_budget\_usd}$)

Claude Agent SDK configuration parameters used to bound long-running loops; exceeding them triggers a ResultMessage with an error subtype ($\text{error\_max\_turns}$ or $\text{error\_max\_budget\_usd}$) that the application can catch deterministically.

Context Isolation

An architectural pattern that delegates noisy, high-volume tasks (such as codebase exploration) to a subagent with its own context window, passing only a concise summary back to the primary agent's context.

Context Pruning and Compaction

A context engineering technique that strips or summarizes verbose tool outputs and historical turns after use, preserving attention focus and token budget for critical reasoning.

Prompt Caching Prefix Matching

A caching mechanism requiring identical, stable content (such as system instructions, tool definitions, and shared knowledge bases) to be placed at the very start of the prompt before dynamic user content in order to achieve cache hits.

Prompt Caching Economics

A billing model where writing a prefix to the cache costs a slight premium over standard input rates, but subsequent cache reads receive a substantial discount upon reuse within its time-to-live.

Message Batches API

An asynchronous Anthropic API endpoint designed for large, latency-tolerant workloads that processes requests within a $24$-hour turnaround window at a significant per-token discount.

Token Streaming

A delivery mechanism that progressively renders tokens to the user interface as they are generated, addressing perceived latency and preventing the user experience from feeling unresponsive.

Headless Mode ($\text{claude -p}$)

A non-interactive execution mode for Claude Code that runs without a human operator at a terminal, using pre-configured permissions to output machine-readable results directly into CI/CD pipelines.

CLAUDE.md Hierarchy

The tiered configuration structure where a repository-committed project-level CLAUDE.md provides shared build commands and standards for all collaborators, while a user-level CLAUDE.md layers individual preferences on top.

Trace Analysis

The debugging discipline of inspecting the retrieved documents, assembled prompt, and raw model generation for a failed interaction to determine whether an error stems from integration retrieval or model reasoning.

Temperature $0$ Sampling

A greedy sampling configuration that maximizes output stability but does not guarantee byte-identical, deterministic outputs across repeated API calls.

Shared Context Window Budget

The architectural reality that a model's context window limit (such as $200{,}000$ tokens) applies to the total sum of input tokens plus the allocated output token reservation ($\text{max\_tokens}$).

Multi-Shot Prompting

The practice of embedding two or three concrete sample input-output pairs inside the prompt to reliably demonstrate and enforce a desired output structure or format.

Adaptive Thinking

A configuration mode where Claude evaluates task complexity to decide whether and how deeply to engage extended thinking, guided by an $\text{effort}$ parameter that trades reasoning depth against token cost.

Claude Haiku

The fast, low-cost model tier engineered specifically for simple, lightweight, and high-volume tasks where tight latency and unit economics are paramount.

Fast Mode ($\text{speed: "fast"}$)

An opt-in configuration utilizing a beta header that serves the same underlying Claude model (such as Opus) with up to $2.5\times$ faster output token generation at premium per-token pricing.

Exponential Backoff with Jitter

A client-side retry strategy that progressively doubles wait times and adds random delays when encountering HTTP $429$ (rate limit) or HTTP $529$ (overloaded) responses to avoid synchronizing retry spikes.

Model Version Pinning

The configuration discipline of targeting specific, dated model snapshot identifiers in production rather than floating aliases, ensuring model upgrades are validated deliberately against an evaluation suite.

Defense-in-Depth (AI Security)

A layered security posture that combines prompt guidance, structural content boundaries, least-privilege tool permissions, human approval gates, input validation, and output filtering rather than relying on prompt text alone.

Model Context Protocol (MCP)

An open standard enabling centralized servers to expose data and tool capabilities consistently to diverse clients, including Claude Desktop, Claude Code, and custom applications.

MCP Transports ($\text{stdio}$ vs. HTTP)

The two primary communication modes in MCP: $\text{stdio}$ for single-machine local process-to-process communication, and network transports (such as streamable HTTP) for remote, shared multi-client services.

Skills vs. Tools

The capability distinction where Skills encapsulate procedural knowledge, checklists, and templates loaded on demand, whereas Tools provide functional schema definitions for invoking external systems.

Curated Toolset Design

The best practice of equipping an agent with a small, focused set of distinct, non-overlapping tools whose outputs return concise, relevant data fields rather than raw, verbose payloads.

Content Boundaries

The structural delimitation of user-supplied data using explicit tags and system instructions, ensuring the model treats external text as inert data rather than executable instructions.

Characterization Tests

Automated tests authored prior to refactoring that lock down and capture the existing behavior of legacy systems, ensuring subsequent modifications do not introduce silent regressions.

Bring Your Own Cloud (BYOC)

A deployment architecture where models or agents run within the customer's dedicated cloud account (e.g., AWS Bedrock) to satisfy strict enterprise data residency and governance policies.