Reading scenario questions carefully

Jules's first practice set goes badly in an instructive way. Of ten original scenario questions Ravi has written, Jules gets six right, and on three of the four misses Jules knew the material perfectly well. One question asked for the option that needed no developers, and Jules chose a custom agent because Birchline had built one. Another asked how to keep customer names out of a model provider's logs, and Jules chose human review, which is good practice but does not answer the question. A third asked which offering would help call centre staff during live calls, and Jules chose a search tool for the website. "You don't have a knowledge problem on those three," Ravi says. "You have a reading problem. That is good news, because reading is the fastest thing to fix." This lesson is about that fix.

The four parts of a scenario

Almost every scenario question on a foundational Google Cloud exam can be split into four parts. The context tells you who the organization is: its size, its industry and what it already has. The need tells you what it wants to achieve. The constraints tell you what limits the choice: no developers, a short deadline, strict privacy rules, a small budget, a need for consistency or a need for variety. And the ask tells you exactly what the question wants: the best option, the first step, the most cost-effective approach, the main risk, or which two statements are true. Reading for those four parts, in that order, is the habit that turns knowledge into marks.

The context and the need usually narrow the field to two or three plausible options. The constraints and the ask then pick the winner. In the law firm example in the figure, several options might sound reasonable: fine-tuning a model on the contracts, writing better prompts, a general chat assistant, or an enterprise search and grounding offering that answers from the firm's documents with citations. The constraint of no AI developers rules out building something custom from scratch. The need for citations points to grounding over fine-tuning, because grounding retrieves the source passage and can show it, while fine-tuning changes the model's behaviour without giving the answer a source. The best answer is the managed grounding option.

Signposts in generative AI questions

Signposts are phrases that point strongly toward one kind of answer. They are not tricks, and they are not a substitute for understanding, but recognizing them quickly saves time and catches misreadings. Ravi asks Jules to start a list in the study log and add to it every time a practice question turns on a phrase. After two weeks the list looks like the table below. Yours will differ, and building it yourself is more useful than memorizing ours.

Signposts work best as a second check, not a first reflex. Read the whole question, decide what you think the answer is, then look for the signpost to confirm it. If the signpost points somewhere else, slow down and reread. Sometimes a question contains two signposts that pull in different directions, such as "the company's own documents" and "a consistent brand voice", and the ask tells you which one matters for this question.

Best fit, not merely works

The hardest questions on the exam are those where more than one option would work. A foundational exam asks for the best fit: the option that meets the need and respects every constraint with the least effort, cost and risk. A custom agent built on the developer platform can do almost anything, so it would work for most scenarios. That does not make it the best answer when a prebuilt tool does the same job for a team with no developers. Fine-tuning can improve almost any output, but it is rarely the best first step when better prompts or grounding would solve the problem faster and more cheaply.

Ravi gives Jules a rule of thumb for these questions: prefer the simplest option that fully meets the need. Simpler usually means prebuilt over custom, prompting over grounding when no outside data is needed, grounding over fine-tuning when facts are the problem, and managed over self-built. When two options seem equally simple, the constraints decide. The rule is not absolute; sometimes the need genuinely requires the more complex option. But it catches the most common mistake: choosing the most powerful answer when the question asked for the most suitable one.

Watch the ask

The ask is the part candidates most often skim, and it is the part that most often decides the answer. "What should the organization do first?" is a different question from "What is the best long-term approach?" A first step is usually about understanding the problem, the data or the risk: defining the business goal, checking data quality, or running a small pilot. A long-term approach may be a full rollout with monitoring. "Which is the main risk?" asks you to rank, not list. "Which statement is not true?" reverses the usual logic, and a candidate who reads quickly will choose a true statement.

Jules's fourth miss in the first set was exactly this. The question asked what a retailer should do first before launching a customer-facing assistant, and Jules chose "deploy it with monitoring dashboards", which is a good later step. The best first step in the options was to define success measures and test the assistant on a set of real customer questions. Jules now underlines the ask in every practice question, in the study log, before reading the options.

# A reading checklist applied to one practice question
question = {
    "context": "online bookshop, 15 staff, no developers",
    "need": "answer customer questions about orders and returns",
    "constraints": ["no developers", "answers must follow the returns policy"],
    "ask": "best option",
}
for part in ("context", "need", "constraints", "ask"):
    print(part.upper() + ":", question[part])
print("Prefer a prebuilt, grounded option that uses the returns policy as its source.")

CONTEXT: online bookshop, 15 staff, no developers
NEED: answer customer questions about orders and returns
CONSTRAINTS: ['no developers', 'answers must follow the returns policy']
ASK: best option
Prefer a prebuilt, grounded option that uses the returns policy as its source.

Eliminating options

When the best answer is not obvious, eliminate. Most questions have one or two options that are clearly wrong: an offering from the wrong audience, a technique that solves a different problem, or a statement that is simply false. Remove those first. Then compare the remaining options against the constraints one at a time. An option that breaks a constraint is out, however attractive it sounds. If two options are still left, ask which one needs less building, less data, less cost or less risk, and choose it unless the need clearly calls for the other.

Be careful with options that are true but irrelevant. "Human review improves quality" is true, but if the question asks how to keep names out of logs, it does not answer it. "Gemini models are multimodal" is true, but if the question asks which offering lets employees draft documents in tools they already use, a fact about the models misses the point. True statements that do not answer the ask are among the most effective distractors on the exam.

Words that change the meaning

Small words matter. "Only", "always" and "never" make a statement easier to falsify, so be wary of options built on them. "Most", "best" and "first" ask you to rank. "Not" and "except" reverse the question. "Prebuilt", "custom", "managed" and "low-code" describe how much building is involved and often line up with the constraints. In generative AI questions, also watch for "factual", "creative", "consistent", "current" and "private", because each points to a different technique. Reading slowly enough to notice these words costs a few seconds per question and saves many marks.

More practice scenarios

Here are three more original scenarios to read with the four parts. First: a regional newspaper wants an assistant that answers readers' questions about local events from this week, and its newsroom has one developer. The context is a small newsroom; the need is answers about current events; the constraint is a single developer; the ask is the best option. A prebuilt offering grounded on current sources fits, because the problem is recency and the team is small. Second: a pharmaceutical company wants to summarize clinical study reports, and a mistake could mislead doctors. The ask is the most important safeguard. Human review by qualified staff is the answer, even though grounding also helps. Third: a game studio wants many unusual names for fantasy creatures. The ask is which setting to change, and the answer is a higher temperature.

Look at how often the ask decides between two good practices. Grounding and human review are both good, but only one is the most important safeguard where errors could mislead doctors. Prebuilt and custom offerings both work, but only one fits a newsroom with one developer. Train yourself to finish each scenario by saying the ask in your own words before you look at the options.

By the end of the week, Jules has reworked the four missed questions and written a fifth of Jules's own for each of them, using the same trap. Writing a question with a deliberate trap is one of the fastest ways to learn to spot one.