2 Claude fundamentals and choosing the right tool · CCAO-F

Exam domain 3: Product and Model Selection (12%)

The right answer in this domain is almost always the cheapest, fastest feature and model that still meets the quality bar. The exam tests whether you can match a task to a Claude feature, a model tier, and a sensible way of managing context.

What Claude is

Claude is a family of large language models made by Anthropic. It reads and writes text, reads images and documents, analyzes data, and can create files. It predicts likely text from patterns. It does not look facts up unless a tool such as web search or a connector gives it sources. That single fact explains most of its strengths and its main risk, hallucination (Chapter 4). You reach Claude through the Claude apps (web, desktop and mobile) on Free, Pro, Max, Team and Enterprise plans. Feature availability and usage limits differ by plan and change often, so the exam focuses on which feature fits a task, not on plan-by-plan limits.

Choosing a feature

Choosing a model tier

Anthropic offers three model tiers. The exam names them Haiku, Sonnet and Opus. Version numbers change; the trade-offs do not.

Tier Strength Typical business use

Haiku Fastest and lowest cost High-volume, simple work: classifying tickets, short replies, extracting fields, summarizing many short items

Exam rule of thumb. Start with the smallest model that meets the quality bar. Move up when output quality falls short, not by default. Anthropic's own sample question rewards choosing a faster, cheaper model for high-volume customer-reply drafts and reserving the top model for complex reasoning.

Some models also offer extended thinking, where Claude reasons step by step before answering. It helps on hard analysis and planning, and costs time and usage. Skip it for simple tasks.

Managing context and memory

The context window is everything Claude can see at once: instructions, the conversation, uploaded files and tool results. It is measured in tokens (a token is roughly three-quarters of an English word). Long chats fill it. As it fills, early details can get less attention and usage limits arrive sooner.

Situation Best move
The chat has drifted onto a new topic Start a new chat
A long chat still matters, but is slowing or losing
early details

Ask Claude for a summary of decisions and open items, then paste it into a new chat

You repeat the same background every time Persist it: Project instructions and knowledge, or memory CLAUDE FUNDAMENTALS AND CHOOSING THE RIGHT TOOL

Situation Best move
The source is a huge document but you need
one section
Upload or paste only the relevant part

Information is sensitive or temporary Use an incognito chat so it is not saved to memory or history

How Claude works, in plain terms

You do not need machine-learning theory for CCAO-F, but four ideas explain most exam scenarios. Claude generates, it does not retrieve. Without a tool, Claude writes the most plausible continuation of your request based on patterns learned in training. That is why it writes well and also why it can invent a statistic that sounds right.

Claude only knows what it can see. In any chat, Claude works from its training plus whatever is in the context window: your messages, Project instructions, uploaded files, memory and tool results. If a fact is not in that window and not in its training, Claude cannot know it.

Training has a cutoff. Claude's built-in knowledge stops at a point in time. Anything newer needs web search, Research mode or a connector.

Tools extend Claude. Web search, Research, connectors, code execution and file creation let Claude look things up, read live business content, calculate exactly and produce real files. Choosing the right tool is half of Domain 3.

Plans and where features live

The exam does not test plan prices or exact limits, but you should understand the general shape:

Individual power users Higher usage, more models and features such as Projects, Research and extended thinking; Max raises limits further

Team Small and mid-size or‐

ganizations

Shared Projects, central billing, admin controls and commercial data terms Enter‐

prise

Large organizations Everything in Team plus stronger security, identity, audit and governance controls, and larger context options

Two takeaways matter on the exam. First, organizational work should run on an approved business plan, not a personal account (Chapter 7). Second, when a scenario asks which feature to use, assume the feature is available unless the question says otherwise.

Worked examples

Worked example 2.1: One task, three possible tools

Situation. Priya, an operations analyst, must answer: "What changed in our three main competitors' delivery policies this quarter?" Weak approach. She asks in a plain chat with no tools. Claude gives a confident answer based on training data that may be a year old. Two of the "changes" it lists are out of date. Better approach. She uses Research mode with a precise brief:

Research the delivery and returns policies of Competitor A, Competitor B and Competitor C as of this quarter. For each, list any change announced in the last 90 days, with the date and a link to the official source. If you cannot find an official source for a change, say so. Present a table: competitor, change, date, source.

Why it works. The question needs current, multi-source, cited information, so Research mode fits. The brief sets a time window, asks for official sources and allows "not found." Priya still opens every link before sharing (Chapter 4).

What the exam is testing. Recognizing that recency plus multiple sources points to Research or web search, not plain chat.

Worked example 2.2: Picking a model tier for a weekly job

Situation. Marcus runs a support team. Every Monday, 1,200 customer survey comments need tagging as Praise, Bug, Feature request, Billing or Other.

Option Result Verdict
Opus with extended thinking Excellent tags; slow and uses us‐
age limits quickly
Overkill for a simple, rules-based
task

Sonnet Excellent tags; moderate speed Fine, but more than the task needs Haiku with clear category definitions and

three examples per category

Very good tags; fast and low cost Best fit, with a 50-comment spotcheck each week Lesson. For high-volume, well-defined work, the smallest tier that meets the bar wins. Quality comes from clear definitions and examples more than from a bigger model. Marcus moves up a tier only if the spot-check shows too many errors.

Worked example 2.3: When to move up a tier

Situation. Elena, a policy advisor, asks Sonnet to compare two 60-page draft regulations and identify where they conflict. The answer misses subtle conflicts in definitions.

CLAUDE FUNDAMENTALS AND CHOOSING THE RIGHT TOOL

Better approach. She switches to Opus, enables extended thinking, and asks Claude to first list every defined term in each draft, then compare definitions, then trace where differing definitions change obligations. Lesson. Complex, multi-step reasoning over long documents is where the most capable tier and extended thinking earn their cost. She also decomposed the task, which helps any model.

Worked example 2.4: Rescuing a long chat

Situation. Over three days, Tom has used one chat to plan a product launch. It now runs to more than 150 messages. Claude has started contradicting decisions made on day one. Fix. Tom sends:

Summarize this conversation for handover. List: (1) final decisions with dates, (2) open questions, (3) owners and deadlines, (4) any constraints we agreed. Use headings. Do not include discarded ideas.

He checks the summary, corrects one item, and starts a new chat inside a Launch Project. He pastes the summary as the first message and adds the launch brief to the Project's knowledge. Lesson. Summarize, restart and persist. The exam often describes this exact symptom and offers "keep going and repeat yourself" as a wrong answer.

Worked example 2.5: Choosing an output surface

Situation. A training coordinator needs a one-page onboarding checklist that managers will reuse and tweak for each new hire. Choice. An Artifact (or a created document file), not inline text. The checklist is substantial, will be edited repeatedly, and will be shared. Asking Claude to "put this in an artifact so I can iterate on it" keeps each revision in one place.

Scenario walkthroughs

Work through each scenario before reading the analysis.

Scenario A. A sales manager wants a quick rewrite of one awkward sentence in an email. Analysis. Plain chat, default model. No Project, no Research, no Artifact. Over-engineering a trivial task is a wrong answer in disguise.

Scenario B. A nonprofit's grants team writes eight to ten grant applications a year. Each draws on the same mission statement, program data, budget templates and past winning applications. Analysis. A Project with the shared materials as knowledge and instructions describing tone, funder expectations and formatting rules. Each application gets its own chat inside the Project. Scenario C. A finance analyst uploads a 40,000-row CSV of transactions and asks for monthly totals by region, with a chart.

Analysis. Code execution and file creation. Claude should calculate with code, not estimate in prose, and can return a chart and a spreadsheet. The analyst spot-checks two months against the source system.

Scenario D. A marketing team lead asks Claude what the latest version of a competitor's pricing page says.

Analysis. Web search, because the answer depends on a live web page. Training data cannot be trusted for current prices.

Scenario E. A consultant wants Claude to remember that they prefer British spelling and bulletpoint summaries across all their chats. Analysis. Memory, or a personal preference setting, because the preference applies everywhere. If the preference applied only to one client, it would belong in that client's Project instructions instead.

Scenario F. An HR partner needs to discuss a sensitive employee situation and does not want it stored in chat history or memory.

Analysis. An incognito chat, on an approved business plan, with personal identifiers removed where possible. Chapter 7 covers the policy side.

Key points

Match the feature to the job: chat for one-offs, Projects for recurring work, Research for multi-source questions, Artifacts for reusable output.

Haiku for volume, Sonnet as default, Opus for hard reasoning.

Restart, summarize or persist. Know which one a scenario calls for.

Common traps

Choosing the most capable model "to be safe" when the scenario stresses volume, speed or cost.

Using a Project for a one-time question, or a plain chat for weekly recurring work. Pushing on in an overloaded chat instead of summarizing and restarting.

CLAUDE FUNDAMENTALS AND CHOOSING THE RIGHT TOOL

Check yourself

A team must summarize 300 short daily sales notes into one-line updates. Speed and cost matter most. Which model tier is the best starting point?

What three options do you have when a long conversation starts losing early details? Name the feature that best fits: "a cited briefing on how five countries regulate electric scooters, as of this month."

Why is a plain chat a poor choice for questions about this week's events?

When is an Artifact a better output surface than inline text?

Answers. (1) Haiku or the fastest, lowest-cost tier, with spot-checks. (2) Restart, summarize-and-restart, or persist the context in a Project or memory. (3) Research mode. (4) Claude's training has a cutoff and plain chat has no live lookup. (5) When the output is substantial and will be edited, reused or shared.

Exercise

List five tasks from your own week. For each, write down the feature and model tier you would choose, and one sentence on why.