Prompt Engineering

Ask a model to "write about our new running shoe" and you get generic marketing copy. Ask it to "write a 40-word product description for a lightweight trail-running shoe, aimed at weekend hikers, in a friendly tone," and the output is usable. Nothing about the model changed between those two requests. Only the wording did, and that is the whole idea of prompt engineering.

A prompt is the text input you send a large language model (LLM). Prompt engineering is the process of crafting that input to reliably get the output you want, described by Google as an iterative, test-driven process of refining inputs to achieve a specific outcome^[1]. You write a prompt, read the response, and adjust the wording, repeating until the answer is accurate and well formatted. The diagram shows that loop.

Why a leader should care

Prompt engineering is the fastest, lowest-cost lever^[2] for improving model output. It needs no training data, no labeled examples, and no infrastructure, so a team can improve results in an afternoon instead of a project cycle. That makes it the first lever to reach for, and often the only one a use case needs.

One lever among siblings

Prompt engineering changes the words you send in. It does not add new knowledge to the model or permanently change how the model behaves. When answers must reflect current or private facts, the right lever is grounding, and when the model must consistently perform a specialized task, the right lever is fine-tuning, which adapts the model's weights on labeled examples. Both are covered in their own subtopics. Keeping these separate matters because the exam rewards matching the lever to the problem, and reaching for a prompt tweak when the real gap is missing facts is a classic wrong answer.

Suppose you need a model to classify incoming support tickets as billing, technical, or account. Depending on how much you show it in the prompt, you are using one of several named techniques, and they form a clear ladder from showing nothing to showing several examples.

Zero-shot, one-shot, few-shot

A zero-shot prompt gives the model an instruction and no examples, relying on its pre-existing knowledge^[1] to respond. A one-shot prompt adds a single worked example of the input and the desired output. A few-shot prompt adds several. Examples teach the model the format, phrasing, and scope you expect, so Google recommends including few-shot examples^[3] in most prompts and notes that prompts without them are often less effective. The counterweight is that if you include too many examples, the model starts copying them too closely instead of generalizing to new inputs, which is what overfitting means, so a few well-chosen examples usually beat a long list. The diagram places the three on that ladder.

Role prompting

Role prompting, also called persona prompting, tells the model who to act as. Casting it as a specific persona^[4] such as a financial analyst, a friendly retail assistant, or a security reviewer steers the vocabulary, tone, and depth of the answer without any examples. A practical rule from Google's guidance: if you assign a role, define that role in the prompt, because a prompt that refers to a persona it never established leaves the model guessing.

Prompt chaining

Prompt chaining breaks a complex job into a sequence of smaller prompts and feeds each output into the next. Rather than asking one prompt to research, outline, and write a report, you chain the prompts^[3]: one summarizes the source, the next drafts an outline from that summary, and a third writes each section. Each step is simpler, easier to check, and easier to fix. The cost is that an error early in the chain flows downstream, so the earliest steps deserve the most scrutiny.

What the three have in common

All three reshape the prompt rather than the model, and each answers a different question: shots decide how much you show the model, role prompting decides who it speaks as, and chaining decides how many steps the job is broken into. Read a scenario for which of those three is missing and the technique names itself.

A single instruction is fine for looking up a fact, but it struggles with a word problem that needs several steps of arithmetic, or with a task that needs information the model does not have. Two advanced techniques address those two different gaps.

Chain-of-thought: show the reasoning

Chain-of-thought (CoT) prompting asks the model to explain its reasoning step by step^[5] before giving a final answer. Google's glossary describes CoT as a technique that has the model detail its intermediate reasoning^[1] rather than jump straight to a conclusion, which significantly improves performance on complex reasoning tasks. Do not confuse it with prompt chaining from the previous section: chain-of-thought asks for the reasoning steps inside a single prompt, while prompt chaining links separate prompts end to end, feeding each output into the next. In practice, adding a line like "explain your reasoning step by step" can turn a wrong one-line answer into a correct worked-through one. The trade is length: the model produces more text, which costs more and takes longer, so reserve CoT for problems that genuinely need multi-step reasoning.

ReAct: reason and act

CoT reasons but cannot look anything up. ReAct, short for reasoning and acting, adds that missing half. The model interleaves a reasoning step with an action, such as calling a tool or running a search, then reads the result and reasons again, repeating until it can answer. The diagram shows that loop of thought, action, and observation. This reasoning-and-acting pattern is the conceptual basis of AI agents, which on Google Cloud run on Vertex AI Agent Engine^[6], a runtime where the model decides which actions to take and in what order. The agent tooling subtopic covers the products. Here the point is that ReAct is what lets a model move beyond its own knowledge and interact with live systems.

Which of the two

Reach for CoT when the task is self-contained reasoning the model can do from what it already knows, such as math, logic, or planning. Reach for ReAct when the task needs facts or actions from outside the model, such as checking an order status or querying a live database. CoT makes the thinking visible; ReAct makes the model able to act on the world.

Exam questions here describe a business situation and a symptom, then ask which prompting technique fixes it. The answer follows from two signals in the stem: how much the model needs to be shown, and whether it needs to reason or reach outside itself.

Read the symptom, name the technique

If the output has the wrong format or tone, the fix is prompt engineering in general, usually few-shot examples^[3] that demonstrate the format. If the stem stresses a consistent structure across many responses, that is few-shot. If it stresses a particular voice, persona, or expertise, that is role prompting. If it describes a long task with several stages, that is prompt chaining. If it involves multi-step math, logic, or reasoning the model keeps getting wrong, that is chain-of-thought. If it needs current data, a lookup, or a tool call, that is ReAct.

Common distractors

Two wrong answers recur. The first offers grounding or RAG (retrieval-augmented generation) for a formatting or tone problem, or fine-tuning for a one-off request. Those are heavier levers for different failures: grounding adds facts and fine-tuning changes behavior, when a prompt tweak was all the scenario needed. The second offers more few-shot examples as a cure-all. Past a handful, extra examples cause overfitting and stop helping, and some symptoms, such as stale facts, no number of examples can fix. Match the lever to the symptom: prompting shapes how the model responds, not what it knows.

Watch the vocabulary

The exam uses Google's terms precisely. Zero-shot means no examples, not a quick or careless prompt. Few-shot means a few examples inside the prompt, not fine-tuning on a few files, which is a separate, weight-changing technique. Chain-of-thought is about showing reasoning, while prompt chaining is about linking separate prompts, and the similar names are easy to confuse.