Claude Certified Developer — Foundations objective flashcards

Question 1

What does objective 1.1, agent architecture, cover on the CCDV-F exam?

Accepted Answer

Nine items on choosing a shape. The deciding question is nearly always whether the order of steps is known before the input arrives: where it is, code can sequence the calls and a planner buys variance for nothing. Learn the named patterns as a set, because several are separated by exactly one fact. Orchestrator-workers and sectioning both fan out and differ only in who fixes the branches. Routing exists for two reasons, accuracy and cost. Evaluator-optimizer needs written criteria and a second pass that measurably helps.

Question 2

What does objective 1.2, agent construction, cover on the CCDV-F exam?

Accepted Answer

Four build surfaces with four different owners of the loop, and items usually describe a constraint that eliminates three of them. A library running inside your process, a hosted product where Anthropic runs both the loop and the container, an interactive terminal, and calling the API while writing the loop yourself are not interchangeable. Bounds are the other half of this objective: turn and spend caps default to no limit, delegated work counts against the same total, and a run stopped by a cap returns no final text.

Question 3

What does objective 1.3, patterns and frameworks, cover on the CCDV-F exam?

Accepted Answer

The wire contract and the mechanisms around it. Append the assistant turn whole, answer every call in one following user message, and pair each result to its call by identifier: three rules that account for most rejected requests here. Past that, a failing tool comes back in band with an error flag rather than escaping as an exception, a paused turn resumes by re-sending unchanged, and the context mechanisms differ by scope. Only one of them survives a process exiting overnight.

Question 4

What does objective 2.1, understanding needs, cover on the CCDV-F exam?

Accepted Answer

Sorting statements before designing against them. A business target leaves every build decision open and cannot be graded; a functional requirement is demonstrable on a single case; a quality attribute is measurable at the service boundary. Two traps recur. A duration naming no clock can be satisfied and breached at the same time, and a monthly budget with no volume behind it fits every design at some volume. Residency and retention rulings belong here too, since they eliminate platforms before any code exists.

Question 5

What does objective 2.2, systems life cycle, cover on the CCDV-F exam?

Accepted Answer

Phase identification, plus the maintenance driver classical systems do not have: the model underneath is withdrawn on somebody else's calendar. Deprecation and retirement are separate states with separate consequences, and only one takes a service down. Items also ask what makes a change safe to ship under pressure, which is a version you can restore rather than a repair authored during an incident. A search for a model identifier returns call sites and registry entries, and those two want opposite edits.

Question 6

What does objective 2.3, claude api mechanics, cover on the CCDV-F exam?

Accepted Answer

The densest cluster of mechanism questions on the paper. Caching matches a prefix of the rendered bytes, and tool definitions render ahead of the system prompt, which between them explain most reports of reads stuck at zero. The asynchronous path is chosen on whether somebody waits for an individual result and never on volume, and its results return unordered, so reconciling them needs the identifier you sent with each request. Streaming changes when bytes arrive and nothing about what they cost.

Question 7

What does objective 2.4, software engineering, cover on the CCDV-F exam?

Accepted Answer

Ordinary engineering, examined where it meets a non-deterministic caller. Independent calls want bounded concurrency rather than a serial loop or an unbounded fan-out. A restructuring with patchy coverage wants current behaviour pinned first, defects included, so that one job is being done at a time. Retry safety belongs to the operation rather than to the transport, which is why a single policy across reads and writes produces duplicated side effects. An idempotency key is minted once for the intent and resent unchanged.

Question 8

What does objective 2.5, application design, cover on the CCDV-F exam?

Accepted Answer

Where the text goes, and who is able to write into that place. A chat product sends its own instructions, tools and settings with every message, so moving a tuned conversation onto the API is a rebuild rather than a paste. Untrusted material needs delimiters and a sentence saying what the delimited region is for. Two jobs sharing one long conversation contaminate each other, and two customers sharing one is an isolation failure. Schemas and descriptions are read by the model, not only by your parser.

Question 9

What does objective 2.6, config management, cover on the CCDV-F exam?

Accepted Answer

Prompts, model identifiers and schemas are configuration, and these items test whether you treat them as such. A prompt edited in a console leaves no diff, no author and nothing to restore. An identifier read from an unset variable behind a silent fallback is how two environments drift apart with nobody noticing. Current identifiers name one specific model rather than pointing at whatever is newest, so appending a date to one usually produces a string the API has never served.

Question 10

What does objective 3.1, claude code operation, cover on the CCDV-F exam?

Accepted Answer

Eight items, and they reward knowing what each surface costs at load time. Instruction files concatenate up the directory tree rather than override, so a personal habit and a project convention both apply inside a checkout. A skill contributes its description until invoked; a path-scoped rule waits for a matching file; an import expands at launch and saves nothing. Know the modes apart as well: one substitutes a classifier for the human reviewer, one denies instead of prompting, and one drops both.

Question 11

What does objective 4.1, debugging and errors, cover on the CCDV-F exam?

Accepted Answer

Seven items on telling failures apart. The first fact does most of the work: refusals, truncations and paused loops all arrive on a successful response, so a handler wrapped around exceptions reports nothing while users file tickets. Read the stop reason before the content. The second is that a rejected request fails identically forever, which turns backoff around it into a queue that never drains. Capture the request identifier at the moment of failure, because a prompt cannot be traced afterwards.

Question 12

What does objective 5.1, llm fundamentals, cover on the CCDV-F exam?

Accepted Answer

How generation behaves, and what that behaviour costs. Prompt and answer share one window, so a call can be refused before anything is produced or run out of room part way through, and those two report different stop reasons. Reasoning is billed as output and drawn from the same ceiling, which is what breaks a budget sized around visible text alone. Nothing here promises repeatability: no sampling setting ever guaranteed identical output, and several of them are now rejected outright.

Question 13

What does objective 5.2, technical fundamentals, cover on the CCDV-F exam?

Accepted Answer

The transport and the client. An SDK is a typed wrapper over the same REST endpoints, so it replaces stream parsing and retry policy while never replacing conversation state, cache placement or idempotency. Streaming is a one-way event stream over an ordinary POST, which rules out steering a response that is already in flight. Response content is a growing union of block types, so reading the first one and taking its text is a bug waiting for the next feature.

Question 14

What does objective 5.3, model selection, cover on the CCDV-F exam?

Accepted Answer

Choosing a tier, and moving between them. Begin at the cheapest model that could plausibly clear the bar and let a measurement decide, rather than beginning at the strongest and asking an eval to justify stepping down. Aggregate accuracy is a weighted average whose weights are the test set's, so a slice where mistakes are expensive can lose while the total wins. Changing tier is a code change, since parameters are gated per model and a body valid on one is refused by another.

Question 15

What does objective 5.4, cost and tokens, cover on the CCDV-F exam?

Accepted Answer

Where the money actually is. Caching pays from roughly the second or third read and loses on a prefix used once, and its discount reaches input only. A cached prefix still occupies the window in full, so it buys price and prefill speed rather than headroom. The asynchronous path is the structural lever for work nobody is waiting on. Forecasts built from counted input under-read the bill, because output carries the higher rate and is the half you cannot count in advance.

Question 16

What does objective 6.1, context engineering, cover on the CCDV-F exam?

Accepted Answer

A transcript grows because the API keeps nothing between calls, and every turn is charged for all the turns before it. The three instruments are not interchangeable. Pruning clears stale tool results, summarising condenses earlier history, and both die with the session, so anything that has to survive until tomorrow belongs in a store the agent writes and reads. Clearing carries its own price, since it rewrites the prefix and the next request pays to write a cache instead of reading one.

Question 17

What does objective 6.2, prompt engineering, cover on the CCDV-F exam?

Accepted Answer

Placement and precision, examined against the instinct to press harder. Current models follow instructions literally, so one that never stated its reach gets a narrow reading, and capitals will not widen it. Supplying the reason alongside the rule generalises a constraint to cases the rule never listed. For long inputs the documented order puts the document first and the question last. An operator instruction arriving mid-conversation goes after the cached history rather than into the prompt at the front.

Question 18

What does objective 6.3, output handling, cover on the CCDV-F exam?

Accepted Answer

What a consumer does with a response. The structural guarantee a schema gives covers field names and types and stops there: numeric ranges are checked on the client, so a value outside its declared bounds is a validation failure to handle rather than evidence of a misconfigured schema. Read the stop reason before reaching for a parser, since truncation, a decline and an unfinished turn each want a different response. Repair loops need machine-generated errors, a cap, and a defined place to land.

Question 19

What does objective 7.1, application security, cover on the CCDV-F exam?

Accepted Answer

Two threat models worth separating: the user as adversary, and a trusted user with content that somebody else wrote. Most items are the second. Third-party text belongs in a tool result, which is the channel the model reads with suspicion, and your own instruction goes in the turn after it rather than inside the payload. Authenticating a connection establishes who is speaking and says nothing about whether what they said carries authority. Paths and commands the model produces are untrusted input to your code.

Question 20

What does objective 7.2, guardrails and rollout, cover on the CCDV-F exam?

Accepted Answer

Where a rule is enforced decides whether it holds. Permission rules evaluate deny, then ask, then allow, and the first match settles the question, so a narrow allow written underneath a broad deny is never reached at all. Begin from denied and opt in, because a configuration that subtracts from everything cannot exclude a capability that did not exist when it was written. Gate on reversibility rather than importance, and note that gating is possible only where the harness can see what is happening.

Question 21

What does objective 7.3, claude hooks, cover on the CCDV-F exam?

Accepted Answer

A small objective with sharp mechanics. The pre and post split is a split in capability: one fires before a tool runs and can stop it or rewrite its input, the other fires afterwards and can log, record and attach context while being unable to change the result on its way into the conversation. Blocking is an exit-code contract, and a status that merely signals failure reads as the callback itself having broken. Something shared through a repository can be switched off locally.

Question 22

What does objective 7.4, secrets and keys, cover on the CCDV-F exam?

Accepted Answer

The model never needs a credential in order to cause a call that carries one. Trusted code attaches it on the way out, and every alternative puts the secret somewhere durable, since a prompt, a message or a tool argument all survive in history, transcripts and traces long after the request finished. A key in a shipped bundle is public, so remediation is rotation now plus a backend rather than obfuscation. Keys belong to the organisation, which is why removing a leaver leaves theirs working.

Question 23

What does objective 8.1, tool implementation, cover on the CCDV-F exam?

Accepted Answer

A tool is a name, a description and a schema, and selection runs on the description. Say when to call it and when not to, because a sentence explaining only what it does gives the model nothing to test an incoming message against. Failures return as results carrying an error flag rather than as exceptions, and an empty result is worse than an error, since the model cannot tell it from an honest answer of nothing. Dependent calls cannot be dispatched together.

Question 24

What does objective 8.2, mcp server development, cover on the CCDV-F exam?

Accepted Answer

Three primitives separated by who drives them: the model calls tools, the application loads resources, a person picks a prompt. Transport follows deployment shape, and the local one carries a framing rule that catches everybody once, because a print statement lands in the middle of the protocol stream. Authenticating a connection establishes which workspace is talking and decides nothing about which tenant a call may read, so that check belongs inside the server, against a caller it has identified.

Question 25

What does objective 8.3, agentic customization, cover on the CCDV-F exam?

Accepted Answer

A sorting objective. Procedural text with nothing executable behind it, and relevance that comes and goes, is what a skill is for, because always-present instructions are paid for on every single turn. A server earns its transport and its authentication where several independent consumers need the same access, and not before. Built-in capability reaches only what Anthropic's infrastructure reaches, which rules it out for anything private. A subagent answers pollution rather than cost, keeping exploration in its own window.

Every Claude Certified Developer — Foundations objective, on one card each

D1Agents and Workflows