G Flashcards for final review · CCDV-F

Cover the right column and answer aloud. Cards are grouped by domain in rough proportion to exam weight: 120 cards in total. Mark any you miss and revisit them daily in the last week.

Agents and Workflows (15 cards)

Question Answer

Workflow vs agent in one line each? Workflow: code orchestrates fixed steps. Agent: the model decides steps and tool calls in a loop.

Four criteria before building an

agent?

Complexity, value, viability, cost of error.

Prompt chaining is for? Fixed sequential subtasks with optional gates between steps. Routing is for? Inputs in distinct categories sent to specialized prompts or models. Two forms of parallelization? Sectioning (independent subtasks) and voting (same task several times). Orchestrator-workers is for? Subtasks that cannot be predicted in advance.

Evaluator-optimizer needs? Clear evaluation criteria and measurable gains from iteration. Three benefits of subagents? Context isolation, parallelism, specialization. Main cost of multi-agent systems? Much higher token use and hand-off information loss. Tool Runner vs Agent SDK? Tool Runner loops over your tools inside the API SDK; Agent SDK is the Claude Code harness with built-in tools.

Who hosts Managed Agents? Anthropic runs the loop and a per-session sandbox. What makes an action deterministic

in an agent?

A hook, not a prompt.

Three frameworks named in the

guide?

Strands, LangGraph, PydanticAI.

Agent loop exit condition? stop_reason other than tool_use, plus an iteration cap. How should subagents hand back

results?

Compact, structured summaries with a fixed format.

Applications and Integration (30 cards)

Question Answer
The single endpoint for conversations? POST /v1/messages
Three required request fields? model, max_tokens, messages
Does the API remember conversa‐

tions?

No. Resend full history every call.

Auth header for the first-party API? x-api-key, plus anthropic-version Six stop_reason values? end_turn, max_tokens, stop_sequence, tool_use, pause_turn, refusal What to do on pause_turn? Send the response back so the turn continues.

What to do on refusal? Handle gracefully, log stop_details, consider a fallback. Four usage fields to log? input_tokens, output_tokens, cache_creation_input_tokens, cache_read_input_tokens

Streaming event order (short)? message_start, content_block_start, deltas, content_block_stop, message_delta, message_stop

Delta type for streamed tool input? input_json_delta

Why stream long generations? Better perceived latency and avoiding HTTP timeouts.

Batch API discount? 50%

Batch size limit? 100,000 requests or 256 MB

Batch expiry? 24 hours; expired requests are not billed.

How long are batch results kept? 29 days
How to match batch results? By custom_id
Four batch result types? succeeded, errored, canceled, expired
Upload once, reuse many times? Files API, reference by file_id
Get exact supporting passages? Citations on document blocks

Image or question first? Image (or document) first, question last Thinking mode on current models? Adaptive thinking, controlled with effort Claude inside your AWS account? Amazon Bedrock (or Claude Platform on AWS)

Claude inside your GCP project? Claude on Google Cloud (Vertex AI)

Retryable status codes? 408, 409, 429, 500, 529 (and connection errors)

404 from the API usually means? Wrong or unavailable model ID

413 means? Request too large

Pin model versions why? Aliases can change behavior without a deploy.

What belongs in version control be‐

sides code?

Prompts, tool schemas, CLAUDE.md, settings, eval sets

Session hygiene in one line? Fresh, isolated context per task and per user.

API vs claude.ai prompt difference? The API has no product system prompt; you supply all context.

Claude Code (10 cards)

Question Answer
Generate a starter CLAUDE.md? /init
Personal memory across all projects? ~/.claude/CLAUDE.md
Shared team settings file? .claude/settings.json
Personal, git-ignored settings? .claude/settings.local.json

Deny vs allow conflict? Deny wins.

Skill loading behavior? Progressive disclosure: name and description first, full content when relevant. Where do subagent definitions live? .claude/agents/*.md

Run non-interactively? claude -p "..." (headless)

Live machine-readable progress? --output-format stream-json

Free context mid-session? /compact (summarize) or /clear (fresh)

Eval, Testing, and Debugging (8 cards)

Question Answer

First debugging question? Is it the integration layer or the model output?

Never retry which errors unchanged? 400-class client errors

How to report a tool failure to Claude? tool_result with is_error: true and a helpful message What does a trace show? Every request, response, tool call, result, token count, and latency Cheapest grading method? Code-based (exact match, schema, tests)

LLM-as-judge must be? Calibrated against human labels, with specific criteria Why hold out a test set? To avoid overfitting prompts to the eval What to log on every error? The request_id

Model Selection and Optimization (18 cards)

Question Answer

What fills the context window? All input plus generated output

Zero-, single-, multi-shot? No examples, one example, several examples Is LLM output deterministic? No; design with validation and evals Effort levels? low, medium, high, xhigh, max

Try before a smaller model? The same model at lower effort
Haiku is best for? High-volume, low-latency tasks
Opus is best for? Complex reasoning, coding, and agents
Sonnet is best for? Balanced quality and cost at scale
Exact token counts? /v1/messages/count_tokens
Cache match type? Exact prefix match
Cache render order? tools, system, messages
Default cache TTL? 5 minutes, refreshed on hit
Cache write cost? 1.25x (5 min) or 2x (1 hour) base input

Appendix g: flashcards for final review

Cache read cost? About 0.1x base input (lower on some newest models)

Max breakpoints? 4

Classic silent cache invalidator? A timestamp early in the system prompt Does lowering max_tokens cut cost on short answers? No

Claude API streaming protocol? Server-sent events over HTTP

Prompt and Context Engineering (14 cards)

Question Answer
Role and stable rules go where? System prompt
Per-request data goes where? User turn, tagged

Long documents and the question? Documents first, question at the end Why XML tags? Separate and label instructions, data, and examples Ideal number of few-shot examples? About three to five, diverse and consistent Positive or negative instructions? Positive: say what to do

Prefill on current models? Not supported; use structured outputs or instructions Context bloat fix? Prune tool output, summaries, context editing Context drift fix? Task or memory file, compaction preserving decisions, subagents Compaction vs context editing? Summarize history vs remove specific old content Guarantee JSON shape? Structured outputs, or strict: true for tool inputs Schema-valid but wrong? Add business-rule validation

Confident output? Verify against sources and recompute numbers

Input sanitization example? Escape stray closing tags in user input

Security and Safety (12 cards)

Question Answer

Is a prompt a security boundary? No

Direct vs indirect injection? From the user vs hidden in retrieved content Core injection defense? Separate untrusted data, least privilege, approvals, hooks Does temperature affect injection? No

Does a bigger model fix injection? No; it can be more susceptible Block a destructive command? PreToolUse hook or deny rule

Where do API keys live? Secrets manager or environment, never code or prompts Separate keys how? Per environment and team, in separate workspaces with spend limits Exposed key response? Revoke, replace, review usage, add secret scanning Who authorizes the end user? Your application and tools, not Claude PII approach? Minimize, redact or tokenize, control retention, isolate tenants API calls from a browser app? Never with your key; proxy through a backend

Tools and MCPs (13 cards)

Question Answer

Three parts of a tool definition? name, description, input_schema Most important part? The description

Parallel tool results go? All in one user message

tool_choice options? auto, none, any, tool (any and tool rejected on newest models) Limit to one call per turn? disable_parallel_tool_use

Server tools examples? Web search, web fetch, code execution, tool search Too many tool definitions? Tool search with defer_loading

MCP architecture? Host runs a client per server
Model-controlled MCP primitive? Tools
Application-controlled primitive? Resources
User-controlled primitive? Prompts
Local vs remote MCP transport? stdio vs Streamable HTTP
Skill vs MCP in five words? Skills teach how; MCP gives access

Appendix g: flashcards for final review

Sources

Exam facts were checked on 29 October 2026 against these pages:

Claude Certified Developer – Foundations Exam Guide v1.0 https://anthropic-partners.skilljar.com/page/partner-certifications (PDF, effective July 2026; a copy is saved in the project's sources folder)

Anthropic Partner Academy: certification FAQ https://anthropic-partners.skilljar.com/page/ faq-certifications

CCDV-F certification page https://anthropic-partners.skilljar.com/claude-certified-developerfoundations-certification Pearson VUE: Anthropic Claude Certification Program https://www.pearsonvue.com/us/en/anthropic.html

Claude blog: four role-based Claude certifications https://claude.com/blog/four-role-basedclaude-certifications (23 July 2026) Anthropic API pricing https://platform.claude.com/docs/en/about-claude/pricing Message Batches API documentation https://platform.claude.com/docs/en/build-with-claude/ batch-processing

Udemy blog: Claude certifications guide https://blog.udemy.com/claude-certifications-guide/ (third party; used only to corroborate the fee)

claudecertificationguide.com https://claudecertificationguide.com/ (independent community resource, not affiliated with Anthropic)

Technical content also draws on Anthropic's Claude API, prompt engineering, prompt caching, tool use, Claude Code, and MCP documentation as of October 2026.