Techniques · 10 min read
8 prompt engineering techniques, with examples
Zero-shot, few-shot, step-by-step reasoning, roles, decomposition, self-critique, structured output and delimiters, each with a copy-ready example.
Prompt engineering techniques are repeatable patterns for getting more reliable answers from language models. None of them are secret tricks; they are ways of giving the model better information or a better process. This guide covers the eight techniques that matter most in everyday work, explains when each one helps, and gives an example you can adapt. If you are new to prompting, start with our framework for writing AI prompts, then come back here.
| Technique | Best for |
|---|---|
| Zero-shot | Clear, common tasks |
| Few-shot | Matching a style, label set or format |
| Step-by-step reasoning | Math, logic, multi-factor decisions |
| Role prompting | Expert vocabulary and priorities |
| Decomposition | Large tasks with several stages |
| Self-critique | Catching errors and unmet requirements |
| Structured output | Results that feed other tools |
| Delimiters and XML tags | Separating instructions from material |
Zero-shot prompting: a clear instruction, no examples
Zero-shot means asking for the result directly without showing any examples. Modern models handle a wide range of tasks this way, so it should be your default. It works best when the task is common (summarize, translate, classify by sentiment) and when you are precise about the output.
Classify the sentiment of each customer review below as Positive, Negative or Mixed. Return one line per review in the format: number - label. 1. Fast delivery, but the box was crushed. 2. Exactly what I needed. Would buy again. 3. Stopped working after two days.
If zero-shot answers are inconsistent (different labels, drifting formats, wrong level of detail), that is the signal to add examples.
Few-shot prompting: teach by showing
Few-shot prompting includes a handful of input/output pairs before the real input. The model infers the pattern: your labels, your tone, your formatting. Two to five examples are usually enough. Make them varied, so the model learns the rule rather than copying one example, and keep them in the exact format you want back.
Turn product features into one-line benefits for busy parents. Feature: Dishwasher-safe lid Benefit: One less thing to scrub after dinner. Feature: Holds 1.5 liters Benefit: Enough soup for the whole family in one go. Feature: Locks with one click Benefit: No spills in the car on the way to practice. Feature: Keeps food hot for 6 hours Benefit:
A common mistake is using examples that are all the same length or topic. The model then assumes length and topic are part of the rule. Vary everything except the thing you want copied.
Step-by-step reasoning (chain of thought)
For problems with several steps, such as calculations, scheduling, comparisons with trade-offs or debugging, asking the model to work through the problem before answering can reduce mistakes. You can ask for visible reasoning or, if you only want a clean answer, ask it to reason first and then present just the conclusion.
A team of 4 people each works 6 focused hours a day. A project needs about 180 focused hours, and one person is on leave for 3 of the working days. Work through this step by step: calculate the available hours per day, account for the leave, and estimate how many working days the project takes. Then give the final answer on its own line, starting with "Answer:".
A note on reasoning models: some current models already think internally before answering, and their providers say explicit "think step by step" instructions are unnecessary for them. With those models, describe the goal and constraints clearly and let them reason on their own; save the step-by-step instruction for standard chat models or for when you need to see the working.
Role prompting: borrow an expert's point of view
A role tells the model which expertise, vocabulary and priorities to apply. "You are a tax accountant" and "You are a startup founder" will review the same budget very differently. The role works best when it is specific and paired with an audience, because expertise only helps if the answer is pitched at the right level.
You are a senior UX researcher who has run usability tests for e-commerce checkouts. Review the checkout steps below and list the five issues most likely to cause abandoned carts, ordered by impact. For each, explain the problem in one sentence and suggest one fix. Write for a small shop owner with no design background. Steps: 1. Cart page with a coupon field at the top 2. Mandatory account creation 3. Shipping costs shown only on the final step 4. Payment form with 14 fields
Roles do not grant knowledge the model does not have, and they do not make facts more reliable. Treat them as a way to set perspective and tone, not as a quality guarantee.
Decomposition: break big tasks into stages
Asking for a complex deliverable in one shot ("write my whole marketing plan") tends to produce shallow coverage of everything. Decomposition splits the job into stages and handles each one separately, either in a single prompt with numbered steps or across several messages, where the output of one step becomes the input of the next (often called prompt chaining).
Message 1: List the 5 main questions a first-time buyer has before choosing a standing desk. One line each, no answers yet. Message 2: For each question, write a 2-sentence answer for a beginner. Keep the numbering. Message 3: Using the questions and answers above, draft a buying-guide outline with an H2 for each question and one comparison table.
Decomposition also lets you review and correct each stage before moving on, which is much easier than fixing a long final answer.
Self-critique: ask the model to check its own work
Self-critique adds a review pass. After a draft, the model compares its output with your requirements and revises. It is particularly useful when you have several constraints that are easy to miss, like word limits, mandatory points or banned phrases.
Write a 120-word product announcement for our new offline mode. Requirements: - Mention that it works on Android and iOS - Include one concrete use case - No exclamation marks - End with a link placeholder [LINK] After writing, check the draft against each requirement, list any that are not met, and output a corrected final version.
Self-critique catches format and requirement errors well. It is less reliable for factual errors, because a model can confidently confirm its own mistakes. For facts, check sources yourself.
Structured output: ask for a precise format
When an answer feeds a spreadsheet, a script or another prompt, ask for a strict structure: a Markdown table with fixed columns, CSV, or JSON with named keys. Describe the schema and say what to do with missing data, so the output stays parseable.
Extract the events from the email below. Return valid JSON only, no commentary, as an array of objects with these keys: "title" (string), "date" (YYYY-MM-DD or null), "location" (string or null). If a field is not mentioned, use null rather than guessing. Email: """ Reminder: team lunch moves to Thursday the 14th at Rosa's. Also, the quarterly review is still pending a date. """
Many APIs now offer a dedicated structured-output or JSON mode that enforces a schema. In a regular chat window, a clear schema plus "valid JSON only" is usually enough. Our image to JSON prompt tool applies the same idea to image descriptions.
Delimiters and XML tags: separate instructions from material
When a prompt mixes instructions with pasted content (an article, a contract, customer messages), the model can confuse the two. Delimiters such as triple quotes, Markdown headings or XML-style tags draw clear boundaries. They also make it easy to refer to parts of the prompt by name. Anthropic explicitly recommends XML tags for Claude, and other models handle them well too.
<instructions> Compare the two policies below. Produce a table with columns: Topic, Policy A, Policy B, Practical difference. Then add a 2-sentence recommendation for a team of 10 remote employees. </instructions> <policy_a> [paste first policy] </policy_a> <policy_b> [paste second policy] </policy_b>
Tag names do not need to follow a special vocabulary; descriptive names work. What matters is consistency: open and close each tag, and refer to the same names in your instructions.
Combining techniques
Real prompts usually mix several techniques: a role, delimited material, a few examples, a structured output and a self-check. You do not have to assemble that by hand every time. The AI prompt generator builds role, context, task, requirements and format sections automatically, and its Expert mode adds an example slot plus a quality-check list. Choose Claude as the target and it wraps each section in XML tags. To see which techniques an existing prompt is missing, run it through the prompt checker, and look up unfamiliar terms in the prompt engineering glossary.