Prompt Engineering Fundamentals

What Prompt Engineering Is

Prompt engineering is the practice of writing instructions that get a model to produce what you actually wanted, consistently rather than occasionally. OpenAI defines it as writing effective instructions for a model such that it consistently generates content that meets your requirements; Google describes the same activity as creating natural language requests that elicit accurate, high quality responses. There is no secret syntax and there are no magic words. It is ordinary writing, aimed at a reader who knows a great deal in general and nothing at all about your situation.

The most useful thing to understand early is that it is iterative. Google states plainly that prompt design can sometimes require a few iterations before you consistently get the response you're looking for, and describes the whole activity as test-driven. Because a model's output is not deterministic, the same prompt can give different answers on different days. Your first prompt is a draft, and treating it as one is what separates people who get good results from people who conclude the tool does not work.

Why It Matters

Every mainstream AI assistant takes the same input: plain language, the one part of the system you fully control. These techniques transfer between tools, because all three major vendors document the same core advice in their own words.

The Parts of a Prompt

Google Cloud's documentation breaks a prompt into components, which turns "write a better prompt" into a checklist. The first two are essential; the rest are optional:

• Objective — what you want achieved, including the overarching goal and not only the immediate request.
• Instructions — the step-by-step directions for the task.
• Context — anything the model must refer to and could not otherwise know.
• Persona — who or what the model is acting as.
• Constraints — what the model must and must not do.
• Response format — the structure you want back: JSON, a table, a bulleted list, 200 words.
• Few-shot examples — sample inputs paired with the output you want for them.
• Recap — a short restatement of the constraints and format at the end of a long prompt.

You will not need all eight every time. Knowing which one you left out is the point.

The Habits That Do Most of the Work

Be clear and specific. Google calls this an effective and efficient way to customize model behavior. Anthropic offers a test worth stealing: show your prompt to a colleague with minimal context and ask them to follow it — if they would be confused, the model will be too.

Give it the material. Include the information the model needs instead of assuming it has it. Attaching the actual document beats describing it, every time.

Say what the output should look like, rather than leaving the model to guess the structure.

Make constraints measurable. "Write a brief summary" is a subjective qualifier; "write a summary of three sentences or less" is a constraint. Google's checklist names this exact substitution.

Say what to do, not what to avoid. Anthropic's guidance is to replace "do not use markdown" with "your response should be composed of smoothly flowing prose paragraphs". Positive framing describes a target; a prohibition only rules out one of many wrong answers.

Separate the instructions from the material with delimiters — Markdown headings or XML-style tags. Google adds the rule that matters most in practice: pick one format and use it consistently within a single prompt.

Ask for one thing at a time. If a prompt asks a model to summarise, extract entities, translate and draft an email in one pass, Google's guidance is that it is trying to accomplish too much and should be split. For multi-step work, chain the prompts so the output of one becomes the input of the next.

A Worked Example

Start with a prompt that reliably disappoints:

Summarise this and make it good.

Walk the checklist. There is no objective beyond "summarise", no context, no audience, no measurable constraint, no format, and "good" has no definition. Rewritten:

You are briefing a manager who has not read the attached report and has two minutes. Using only the attached text, write a summary of five bullet points or fewer. Each bullet should be one sentence. Lead with the decision being asked for. If the report does not state a decision, say so rather than inferring one.

Nothing clever happened. A persona set the register, the context was bounded, "good" became five bullets of one sentence each, an ordering rule replaced a hope, and the last sentence told the model what to do when its assumption fails — the step almost everyone skips.

The Order Things Go In

When your prompt contains a long document, order matters. Anthropic's guidance is specific:

1. Put the long document or data at the top, above your query, instructions and examples.
2. Wrap each document so it is clearly separated from your instructions.
3. Put the actual question at the end. Anthropic reports that queries at the end can improve response quality by up to 30 percent in tests, especially with complex, multidocument inputs.
4. Ask the model to quote the relevant parts first, then answer from those quotes.

Google makes the general version of the point: the order of content in a prompt can affect the response, so reordering is worth trying when one underperforms.

Common Misconceptions

• "Shorter prompts are better prompts." Length is not the variable; completeness is.
• "There are magic words." Google's checklist says to remove language that tries to influence performance through emotional appeals, flattery or artificial pressure, and states that model performance will no longer improve and in many cases will get worse.
• "Examples are always required." Examples show a format rather than describe it, but Google warns that too many can make a model overfit, and Anthropic warns that examples which are too alike teach unintended patterns.
• "Asking for step-by-step reasoning always improves the answer." It often helps, but it costs length and latency, and Anthropic notes that a general instruction often beats a hand-written step-by-step plan.
• "A prompt that works today will work everywhere." OpenAI is explicit that different model types need to be prompted differently, and that even different snapshots within one family can produce different results.
• "If the output is wrong, the prompt needs more words." Often it needs fewer and better ones. Google's checklist puts typos, undefined jargon, contradictions and redundant restatement at the top of the list.
• "Prompting can fix anything." Anthropic notes that not every failure is best solved by prompt engineering — some are better solved by changing the model.

Verify What Matters

A confident tone is not evidence of correctness. Where a claim is recent, obscure or consequential, either give the model the source material or use a tool that grounds it in one — Google's guidance is that search grounding should be enabled whenever a model may need to know obscure or recent facts. Asking a model to check its answer against stated criteria catches real errors cheaply, and is not a substitute for checking what matters. And text you paste into a prompt is instructions as far as the model is concerned: Google's checklist flags untrusted user input inserted into a prompt as a major security risk.

What the Questions Cover

This topic's bank is built to the blueprint shown on this page: 40 percent easy, 40 percent medium and 20 percent hard. Easy items check that you hold a concept correctly. Medium items put you in a situation and ask what to do. Hard items give you two defensible options and a constraint. Roughly a third of the bank is scenario-based.

Where to Go Next

Advanced Prompting Techniques goes deeper on chain-of-thought, few-shot patterns and structured outputs, and Prompt Engineering for Code covers what changes when the output is a program. For the same craft inside one product, see Claude Prompting Best Practices, ChatGPT Basics or Gemini Basics. Claude AI Basics covers using an assistant at all, and AI Agents Basics covers what changes when one is given tools.

Sources

All read 2026-08-09. Anthropic — Prompting best practices · Google — Prompt design strategies and Overview of prompting strategies · OpenAI — Prompt engineering