Domain 3 – Product and Model Selection (12%)

• Choose among chat, Projects, research mode, artifacts
• Differentiate Haiku (fast/cheap), Sonnet (balanced), Opus (highest capability)
• Align model choice with task requirements (speed, cost, quality)
• Manage context limits and memory (when to start a new chat, summarise, or persist)

CCAO-F Domain 3 · Task 3.1

What You Need to Know

Four ways to handle work. Match feature to the shape of the task, not habit.

Feature	Dominant signal
Ordinary chat	Simple one-off (single question, single draft, analysis of a document attached in the moment)
Projects	Recurring shared context (same background documents, standing preferences, working material across many conversations)
Research	Cited multi-source synthesis (whole-field survey, how-others-do-it brief, audience who will want to see sources)
Artifacts	Work with its own life (significant, self-contained content likely to be edited, iterated on or reused outside the conversation)

Key Selection Rules

• Ordinary chat is the legitimate default more often than enthusiasts admit
• Re-pasting the same context into chat after chat is Project-shaped work done by hand
• Research is not a quality upgrade for ordinary questions answerable from material you already hold
• An artifact is a container for one evolving output; it does not brief new conversations with standing context

Key Concept: Find the dominant signal first, then match it.

Exam Traps

• Treating an artifact as the place to keep standing context for new chats
• Running everything through a Project or Research to maximise quality
• Treating re-pasting the same context into every chat as diligence

Practice Scenario

Parish council clerk has four pieces of work: monthly agenda-and-minutes cycle needing standing orders, member list and format conventions; one-off briefing on how five nearby councils run allotment waiting lists with sources; single resident’s question about a grant deadline; emergency-contacts sheet revised every quarter and shared. Which plan matches feature to work?

Build Exercise: Audit Your Feature Choices (20 min)

CCAO-F Domain 3 · Task 3.2

What You Need to Know

Three model families form a real trade-off of speed and cost against depth:

Tier	Role
Haiku	Fastest, lowest-cost; high-volume, latency-sensitive, routine tasks
Sonnet	Best combination of speed and intelligence; balanced everyday default
Opus	Most capable reasoning; complex, high-stakes work

No tier is “the good one”. Each is the right choice for work that matches its strengths.

Reading the Signals

• Lighter-tier signals: high volume, routine/repetitive, simple transformation (reformat, extract, standardise), speed matters, individually low-stakes
• Most-capable signals: ambiguity, contradiction to untangle, judgement with consequences, complex synthesis, wrong subtle conclusion costs more than the extra time
• When neither dominates → balanced middle

Size does not decide; the kind of work does. Standardising a long list is lighter-tier work however many rows. Reconciling two disagreeing reports is capable-tier work however short the final page.

The Model Picker

Choose in the chat interface; change at any point; applies from the next response. Routine ground-clearing can run on a lighter tier and the hard synthesis on the most capable, inside one piece of work.

Exam Traps

• Choosing the most capable model for everything because the work is important
• Sending work that needs depth to the fastest tier because it is “just writing”
• Accepting a claim that the lighter tier gives the same quality on demanding work

Practice Scenario

Community energy co-op administrator has two jobs: standardising 600 member records into one address format, and drafting the co-op response to a complex regulatory consultation where three board members gave partly contradictory steers. Keep spending on the routine job as low as possible without risking the quality of the co-op’s position. How should she assign models?

Build Exercise: Match a Week of Tasks to Tiers (25 min)

CCAO-F Domain 3 · Task 3.3

What You Need to Know

Every choice between tiers spends three currencies: cost (usage allowance), time (wait per response), and depth (quality on demanding work). Name the currency a task needs, then pay only that.

Three Questions Name the Currency

1. What does being wrong cost here?
2. How many times will this run?
3. Who is waiting for the answer?

Volume Multiplies Cost and Speed

Run once, the tier difference is seconds and a sliver of allowance. Run five hundred times, it decides whether the job finishes today and whether the allowance survives the week. High-volume routine work belongs on the lighter tier with a sampled spot-check.

Stakes Multiply Quality

Some single outputs feed a consequential decision and need depth. The seconds a lighter tier saves are worthless against the cost of a subtly wrong conclusion. Quality requirements follow the cost of being wrong, not the prestige of the task. High-stakes output then goes to qualified human review, whatever model drafted it.

Split Mixed Workloads

Routine bulk → fast tier + sampled review. Hard core → capable tier + proper review. The model picker makes the split cheap.

Key Concept: Volume multiplies cost and speed; stakes multiply quality. When a stem mixes both shapes, the keyed answer almost always splits the workload.

Exam Traps

• Justifying the fast tier on a high-stakes output because it is “only one letter”
• Justifying the top tier on routine bulk because the work is important
• Escalating to a heavier model when the real defect is a vague brief

Practice Scenario

Regional athletics association preparing annual awards night: 500 personalised invitation letters that vary only in name/club/event, and one difficult letter to a club being suspended for a safeguarding failure where tone, precision and legal-adjacent care all matter. How should the administrator run the two jobs?

Build Exercise: Price a Mixed Workload in Three Currencies (25 min)

CCAO-F Domain 3 · Task 3.4

What You Need to Know

A conversation holds a finite working space. When it reaches its length limit, start a new conversation or use a Project. Treat a long chat like a long meeting: know it will end, and decide what must survive it.

The Three Moves

1. Restart — when work is done, when the chat holds dead ends and superseded drafts, or when it nears its limits. A clean start with exactly the right context beats a marathon chat.
2. Summarise — have Claude summarise decisions, current state and what remains. Verify it yourself against source records (the summary inherits every uncorrected error). The verified summary becomes the new chat’s opening brief.
3. Persist — paste, re-attach, or store in Project knowledge any specific working output the next conversation will need. A standing document that keeps being revised and shared has a fourth home: an artifact.

What Memory Carries, and What It Does Not

Memory is a set of individual topics: your role, projects, communication preferences, ongoing work context. It is context about you, not transcripts, drafts or attached documents. Each Project has its own memory. Memory does not reliably transport an approved outline or a data file into the next chat. Explicit persistence remains the dependable pattern for a specific working output.

Key Concept: “A new conversation would lose all our context” is the continuity illusion. A verified summary and persisted documents carry forward the context worth keeping, at a fraction of the length and the usage.

Exam Traps

• Relying on memory or chat search to carry a specific working output into a new chat
• Pasting an unverified summary into the new conversation as the efficient handover
• Keeping everything in one enormous chat because a restart would lose the context

Practice Scenario

Wedding coordinator has run one enormous conversation across three months of planning. Final month starts now; chat is thick with superseded detail; platform length limits are in sight. Some names and dates were corrected more than once. Best plan?

Build Exercise: Run the Handover Discipline on Your Longest Chat (30 min)