Product & Model Selection

Twelve percent of the exam, around seven items, but it sits underneath every other domain: a badly chosen surface or tier makes prompting, validation and troubleshooting all harder than they needed to be.

What it tests is matching the shape of the work to the shape of the tool. A one-off question belongs in a plain conversation. Work that repeats week after week, with the same background material and the same house style, belongs in a project that carries that context for everyone who uses it. Work that must be answerable from specific files belongs in a session with those documents attached. A question needing many sources gathered and weighed belongs in a longer research-style task, where you trade minutes for depth.

Model tier follows the same logic. Haiku for high volume and simple, well-defined work. Sonnet as the balanced default for everyday business tasks. Opus when the reasoning is genuinely hard or being wrong is expensive.

The idea that unlocks the domain: choose by what the task needs, not by habit. Reaching for the deepest tier by reflex is exactly as wrong as reaching for the cheapest.

• Start from the task, not the tool. Decide what the work actually requires before deciding where to do it.

• One-off question with nothing reusable — a plain conversation. Anything heavier is setup you never earn back.

• Same background every week, or a team working from it — a project, which carries shared material and standing instructions into every chat inside it.

• Answers must come from specific files — attach the documents and work from them, so the response is grounded in the text rather than recalled from memory.

• Many sources to gather, compare and cite — a longer research-style task, and accept that it takes minutes rather than seconds.

• The deliverable is a thing someone opens, edits or forwards — produce a shareable artifact, not a long stretch of chat.

• Haiku when volume and speed dominate: classification, extraction, routine formatting, simple well-defined steps.

• Sonnet as the balanced default for ordinary business work — drafting, analysis, summarizing, most everyday judgement.

• Opus for genuinely hard reasoning, real ambiguity, or high consequence where a wrong answer is expensive.

• A deeper tier costs more and answers more slowly. Pay that only when the extra depth would change the outcome.

• A context limit is a limit on what can be in view at once. A long document may need splitting, section-by-section work, or rolling summaries carried forward.

• Match the surface to the audience. A colleague can read a thread; a client or an executive should receive a finished, self-contained document.

• Defaulting to the deepest tier to be safe. Overspending and slow delivery are real failures, not neutral ones. Habit is never the justification the exam wants.

• Defaulting to the cheapest tier to save money on work that genuinely needs judgement. One bad answer costs more than the saving.

• Spinning up a project for a single question. Structure you never reuse is pure overhead.

• Leaving durable, shared context buried in one person's chat history where no colleague can find or reuse it.

• Pasting a long document into the message instead of attaching it and working from the file.

• Assuming more context is always better. Filling the working view with material that is not relevant degrades the answer rather than improving it.

• Treating does not fit as a dead end. Split it, work section by section, or summarize as you go and carry the summaries forward.

• Launching a long research task when you already know the single source. You pay minutes of waiting for nothing.

• Sending a chat transcript to a client because the content was correct. Format is part of the deliverable, not decoration.

• Changing tier to fix a vague prompt. Missing context and unclear instructions fail at every tier equally.

• Choosing by whatever worked last time. The right tier is a property of this task, not of your routine.

• Ignoring speed as a requirement. If a person is waiting live, the slower, smarter tier is still the wrong answer.

MATCH — match the task to the tool, in this order.

• M is for Material. What has to be in view? Nothing reusable means a plain conversation. The same background every week, or a team sharing it, means a project. Specific files mean attaching the documents. Many sources to gather and weigh means a longer research-style task.

• A is for Audience. Who receives this? A colleague can read a thread; a client or an executive should get a self-contained, shareable artifact in their format.

• T is for Task difficulty. Is the work simple and repetitive, everyday business judgement, or genuinely hard reasoning where being wrong is expensive?

• C is for Choose the tier. Haiku for volume and simple work, Sonnet as the balanced default, Opus only when the difficulty you just named earns it.

• H is for Headroom and haste. Does the material fit in one working view, or must it be split and summarized as you go? And how long will someone wait? A deeper tier costs more and answers slower.

Material and audience settle the setup and the format. Task difficulty settles the tier. Headroom and haste are the sanity check that stops you paying for depth the task never needed.