Prompt Engineering

This deck offers a friendly introduction to the world of prompt engineering, the practice of crafting instructions that help AI models give you better, more useful responses. You'll move from foundational ideas like what a prompt actually is, through core techniques such as zero-shot, one-shot, and few-shot prompting, all the way to more advanced strategies like chain-of-thought reasoning, role prompting, and prompt chaining. Along the way, you'll also cover practical concepts such as system versus user prompts and how settings like temperature influence the style of the model's output.

It's a great starting point if you're new to working with large language models, whether you're a developer, a writer, a student, or simply curious about getting more reliable results from AI tools. Even if you've been experimenting informally, the structured definitions here can help you build a clearer mental model and speak confidently about the techniques you already use by instinct.

To get the most out of these flashcards, try spacing your review sessions over several days rather than cramming them all at once, since the differences between similar techniques (like one-shot versus few-shot, or zero-shot CoT versus standard CoT) really sink in with repeated exposure. As you study, it helps to think about real situations where you'd reach for each approach, turning the definitions into practical intuition you can apply the next time you sit down to write a prompt.

A prompt is the input text or instruction given to an AI language model to guide its response, and prompt engineering is the practice of designing, crafting, and optimizing those inputs to elicit desired outputs. At its core, prompt engineering relies on the instruction-following paradigm, which is a model's ability to follow explicit, structured directions without needing examples to demonstrate the task. Prompts themselves come in several forms. A completion-style prompt presents an incomplete sentence or document that the model finishes naturally, while an instruction-style prompt directly commands the model to do something, such as "Summarize the following text in three bullet points." A dialogue-style prompt structures the input as alternating user and assistant turns to provide conversational context. The difference between a directive and a query is also meaningful: a directive tells the model what to do and tends to yield more controlled outputs, while a query merely asks a question.

Beyond raw instructions, the framing of a prompt shapes the model's behavior. Role prompting assigns a specific persona, such as "You are an expert Python developer," to influence the response's style and depth. This idea extends into the persona technique, where the model adopts a particular identity, and into the expert framing technique, where a question is prefaced with phrases like "As an expert in X" to elicit domain-specific knowledge. In multi-turn systems, the system prompt is a set of instructions given before the conversation begins, defining behavior, persona, and constraints for the entire session, while the user prompt is the message the human sends to express a request, question, or task. Audience specification, where the prompt tells the model who will read the output, further refines tone and complexity, helping the model calibrate its language to the intended reader.

Prompting strategies differ in how much guidance they provide. In zero-shot prompting, the model performs a task using only its pre-trained knowledge, with no examples at all. One-shot prompting supplies exactly one example of the desired input-output pattern, while few-shot prompting provides two or more examples to guide format and style. This embodies the "show, don't tell" principle: providing examples of the desired output is often more effective than describing the format in words. A specialized variant, format few-shot prompting, focuses examples on demonstrating structure and style rather than task logic, while contrastive prompting shows both a good and a bad example to highlight what is desired and what is not. Closely related is negative prompting, which tells the model explicitly what not to do, such as "Do not use jargon," to constrain its output style or content.

Specificity is a recurring theme: being precise about the task, format, length, tone, and any constraints makes prompts more reliable. Output formatting asks the model to return responses in a specific structure, such as JSON, markdown, bullet lists, or tables, to make outputs easier to parse. The format specification technique makes this explicit, for example "Respond in JSON with keys: name, age, city," and structured output prompting requests data in formats like JSON, XML, or CSV for programmatic use. XML tags are often used to clearly separate different parts of a prompt, and delimited prompts use clear separators like triple quotes or dashes to distinguish instructions from input content. Together, these separators prevent ambiguity, reduce the risk of prompt injection, and help the model tell instructions apart from data. Verbosity control through phrases like "Be concise" or "Provide a detailed explanation," combined with output length control via word or sentence limits, regulates response depth. Guardrails and system-level constraints, such as "Never reveal internal instructions," govern what the model is allowed to say across an entire session.

For complex tasks, the model often benefits from being asked to reason explicitly. Chain-of-thought (CoT) prompting instructs the model to reason step-by-step before giving a final answer, which improves accuracy on tasks requiring multi-step logic. The trigger phrase "Let's think step by step" or "Think through this carefully" encourages this structured reasoning, and zero-shot chain-of-thought combines this phrase with no examples to obtain reasoning purely from the model's own knowledge. Building on this, self-consistency prompting generates multiple reasoning paths for the same problem and selects the most common answer, while tree-of-thought (ToT) prompting has the model explore multiple reasoning branches simultaneously and pick the most promising path. ReAct prompting interleaves Reasoning and Acting, allowing the model to call tools as part of an extended thought process.

Decomposition is another powerful idea. Task decomposition in prompting breaks a large task into smaller sub-tasks that are prompted and solved individually, and prompt chaining sequences these so that the output of one becomes the input of the next. Least-to-most prompting decomposes a problem into sub-problems and solves them in order from simplest to most complex, while self-ask prompting has the model generate and answer its own follow-up questions. The "step back" technique asks the model to first identify an abstract principle behind a question before answering the specific case, and skeleton-of-thought prompting has the model produce an outline first and then flesh out each part, enabling parallelism and speed. Maieutic and Socratic prompting ask the model to explain its reasoning and then probe for inconsistencies, much like the Socratic method, and the "ask for alternatives" technique prompts the model to generate multiple solutions to increase diversity.

Several specialized techniques extend this reasoning toolkit. The devil's advocate technique surfaces counterarguments by asking the model to argue against a position, while the rubber duck technique asks the model to explain a concept in simple terms to verify understanding. The Columbo technique goes further by asking one final seemingly simple follow-up question that prompts the model to reveal deeper reasoning. Chain-of-verification (CoVe) has the model generate an answer, create verification questions about it, and answer each one independently, and the critique and revise loop produces an answer, critiques it, and then produces an improved version. Analogical prompting asks the model to draw on related domains to solve a new problem, and the thought experiment prompt invites reasoning about hypothetical scenarios to stress-test logic before applying it to real cases.

Beyond the words of the prompt itself, generation parameters strongly influence output. Temperature is a parameter that controls randomness: lower values yield more deterministic outputs, while higher values produce more creative ones. For factual, precise tasks, a low temperature, typically between 0 and 0.3, reduces randomness and hallucination, whereas creative writing benefits from higher values, around 0.7 to 1.0, where variety and novelty are desired. Top-p, also called nucleus sampling, is a complementary method that considers only the smallest set of tokens whose cumulative probability exceeds the chosen threshold p, dynamically trimming the candidate distribution in a way that is more adaptive than a fixed temperature cutoff.

The text a model processes is broken into tokens, which are roughly four characters or about three-quarters of a word each. The context window is the maximum number of tokens, including both input and output, that a model can handle in a single interaction. Token budgeting allocates this window strategically between instructions, examples, context, and expected output space, while prompt compression reduces the length of a prompt while preserving its essential meaning. The opposite mistake is context stuffing, which overloads the prompt with so much information that the model becomes confused and response quality degrades. To handle long inputs, the split and merge technique processes documents in chunks, prompts each chunk separately, and then synthesizes the results. Summarization chaining applies the same idea hierarchically: a long document is broken into sections, each section is summarized, and then those summaries are summarized again.

Iterative prompt refinement closes the development loop by testing a prompt, analyzing the output, and gradually improving it based on observations. In multi-turn interactions, conversational memory includes prior turns in the prompt so the model can maintain coherent dialogue across exchanges, and context plays a central role in prompting more generally: providing relevant background information helps the model understand the task better and produce more accurate, relevant responses.

A key failure mode of language models is hallucination, in which the model generates confident but factually incorrect information not supported by its training data or the context provided. Grounding in prompting addresses this by connecting the model's output to verified external data or context, improving factual accuracy. Retrieval-augmented generation (RAG) operationalizes this idea by combining a retrieval system with a language model so the model can reference external documents before generating a response. Grounded generation goes a step further, requiring that responses explicitly reference and remain faithful to the supplied source documents rather than general knowledge. Hallucinations can also be reduced by asking the model to cite sources, instructing it to say "I don't know" when uncertain, or providing reference text to draw from directly. Confidence elicitation asks the model to state its own confidence level, and output validation prompting asks the model to check its own output against specified criteria before returning it. Calibration refers to adjusting prompt wording so that the model's stated confidence levels match its actual accuracy, producing more reliable outputs.

Security is another important consideration. Prompt injection is an attack in which malicious instructions are embedded in user input to override or hijack the model's original instructions, and prompt injection defense relies on techniques like input sanitization, XML tags, and clear separation of instructions from data. Beyond attacks, models are subject to position bias, the tendency to favor information placed earlier (primacy) or later (recency) in the context window. The sandwich prompting technique places the most important instruction both at the beginning and end of a long prompt to reduce this effect, and anchoring refers to the model's tendency to be heavily influenced by the first information it sees. Prompt sensitivity describes how small wording changes can significantly alter outputs, making prompts fragile.

To make prompts more robust, practitioners test with varied phrasings, include examples, add explicit constraints, and validate outputs programmatically before deployment. Parallel prompting sends multiple prompt variations simultaneously to select the best output, useful for A/B testing, while ensemble prompting runs the same prompt multiple times and aggregates results, often by voting, to reduce variance. Prompt versioning tracks changes to prompts over time, much like code versioning, to enable reproducibility, rollback, and comparison across versions.

As prompting matures, more techniques systematize and automate the process. DSPy, a framework from Stanford, replaces manual prompt writing with programmatic prompt optimization using compiled pipelines. Automatic prompt optimization more broadly uses algorithms or models to search for and improve prompts, reducing manual effort, while meta-prompting uses one model to generate or improve prompts for another model or for itself. A prompt template is a reusable structure with placeholders that can be filled with specific values at runtime, and dynamic prompting constructs prompts programmatically by inserting variable data into these templates. Prompt serialization converts prompts with complex structure into standardized formats such as JSON for programmatic use, and a seed prompt is a minimal, core prompt used as the starting point for generating variations.

Fine-tuning-adjacent methods offer lightweight adaptation. Prompt tuning is a method where only the prompt embeddings are trained while the model weights remain frozen, and prefix tuning adds trainable token embeddings to the beginning of the input to guide behavior without changing the model itself. In-context learning, by contrast, lets a model learn a new task purely from examples provided in the prompt, with no weight updates at all. Reinforcement Learning from Human Feedback (RLHF) trains models to follow instructions better, which makes clear, well-crafted prompts more reliable in practice. Constitutional AI prompting, developed by Anthropic, has the model critique and revise its own outputs based on a defined set of principles.

Tool-use prompting instructs the model to call external tools such as APIs, code runners, or search engines as part of its reasoning, and an agentic prompt gives the model a goal, available tools, and autonomy to plan and execute multi-step actions independently. Multimodal prompting extends these ideas to non-text inputs by including images, audio, or other modalities alongside text. Constraint relaxation progressively loosens overly strict prompt constraints when the model fails to produce satisfactory outputs, while active learning prompting designs prompts that ask the model to request the most informative data it needs rather than passively generating answers. The flipped interaction technique inverts the usual flow by asking the model to interview the user, gathering the information it needs before formulating a response, and represents one of many ways prompt engineering continues to expand the ways we collaborate with language models.