CCAO-F Glossary - Product and Model Selection

Quick-lookup definitions for every domain, with exam context and links back to the lesson that covers each term.

Model Tier (Haiku / Sonnet / Opus): Claude is offered as a family of distinct model tiers rather than one model with a single dial. Haiku is the fastest and cheapest tier, tuned for high-volume, simple, well-defined work. Sonnet is the balanced mid-tier and the right default for most everyday business reasoning. Opus is the most capable tier, built for long, high-autonomy runs or single highest-stakes decisions, at greater cost and latency per request.; Exam context: The exam tests matching a scenario's volume, complexity, stakes, and autonomy length to the tier built for that shape of work. Common distractors default to the cheapest tier 'to save cost' on long, high-autonomy work, or reach for Opus 'to be safe' on high-volume, simple work.; See also: 3.1 Choosing a Model for the Task
Effort Parameter: A setting that tunes how thoroughly Claude works on a given request — how much it checks its own reasoning and how exhaustively it explores a problem before answering — trading that thoroughness against token cost and response latency, within a single already-chosen model. Anthropic recommends the xhigh effort level for advanced coding and high-autonomy agentic work on Claude Opus 4.8; routine, low-stakes requests are better served by a lower effort level.; Exam context: The exam tests distinguishing effort (how hard the chosen model works on this request) from model tier selection (which model family handles the task at all). A common distractor reaches for a different, more capable model when the actual need is more thoroughness from the model already chosen.; See also: 3.2 Tuning Thoroughness with the Effort Parameter
Artifact: A separate panel in Claude.ai for long-form, self-contained, or iteratively-edited content that benefits from living outside the back-and-forth chat flow — a report, essay, case study, or interactive prototype the user plans to keep revising rather than read once. For interactive content, the panel keeps a running version visible and playable while accepting live edits.; Exam context: Length alone isn't the trigger for an Artifact — self-containment and the intent to iterate are. A common distractor offers 'put it in an Artifact' for a short answer meant to be read once with no iteration need.; See also: 3.3 Choosing the Right Claude Product Feature
Claude Project: A Claude.ai feature that adds persistent context — custom instructions and a knowledge base of reference material — available across many separate future conversations, not just the current thread. A style guide, a client's background, or a recurring set of product facts belongs here instead of being re-pasted into every new chat.; Exam context: The exam tests the 'would otherwise need re-pasting into every new chat' signal for Project-level persistence. A common distractor offers a Project for genuinely one-off information with no future reuse.; See also: 3.3 Choosing the Right Claude Product Feature
Research Mode: A Claude.ai capability that has Claude actively search and synthesize across multiple external sources over several steps, producing a sourced answer grounded in current information rather than only what it already knows from training.; Exam context: The exam tests recognizing when a request genuinely needs current, multi-source, cited information versus when Claude's existing knowledge already covers it. A common distractor offers research mode for a question that doesn't require live, multi-source external sourcing, adding latency for no benefit.; See also: 3.3 Choosing the Right Claude Product Feature
Context Window: The finite amount of prior conversation, pasted material, and instructions Claude can hold onto at once. As a single conversation grows very long, Claude's grip on the earliest content can loosen even though it's technically still part of the thread, showing up as forgotten formatting rules, contradicted earlier decisions, re-asked questions, or imprecise handling of a document pasted many messages earlier.; Exam context: The exam tests recognizing degradation signals and choosing the right fix — restart, summarize, or persist — rather than repeating an instruction more forcefully, which treats a structural problem as a wording problem.; See also: 3.4 Context Limits and Memory
Structured Output: Output such as a table rendered directly in a reply, suited to data meant to be scanned and compared side by side rather than read as prose. It can live inline in chat or inside an Artifact depending on its length.; Exam context: The exam tests recognizing a fundamentally comparative or tabular request as its own signal within the product-feature decision framework, distinct from persistence (Project), current sourcing (research mode), or long-form iteration (Artifact).; See also: 3.3 Choosing the Right Claude Product Feature
Context Degradation: The observable loosening of Claude's grip on earlier conversation content as a thread grows very long — re-explaining something already settled, applying an early instruction inconsistently, re-asking something already answered, or getting details wrong about material pasted much earlier in the same conversation.; Exam context: The exam tests choosing the correct one of three responses — restart, summarize-then-continue, or persist to Project knowledge — based on whether the earlier content is still needed. A common distractor is repeating the instruction more forcefully, which doesn't address the structural cause.; See also: 3.4 Context Limits and Memory
Chunking (Long Documents): Splitting a very long document into manageable pieces — pasting only the relevant section, or requesting a section-by-section summary pass first — instead of loading the whole thing at once. A single very long document competes for the same limited context-window attention as everything else already in a conversation.; Exam context: The exam tests recognizing that context limits affect single long documents, not only long conversations. A 300-page report pasted whole when only two chapters are relevant is the textbook signal for chunking rather than pasting the entire document.; See also: 3.4 Context Limits and Memory