Prompt engineering
Prompt engineering is the study and construction of inputs that condition the behavior of generative artificial-intelligence systems. The field concentrates principally on models whose behavior can be directed through natural-language instructions, contextual demonstrations, structured data, or combinations of these elements. Its central object, the prompt, functions simultaneously as a user request, a temporary specification, and a portion of the computational context from which a model predicts an output.
Prompt engineering became a distinct technical practice with the spread of large language models capable of in-context learning. Unlike conventional software, these systems do not require every task to be expressed through a formal programming language. They nevertheless remain sensitive to wording, ordering, formatting, and contextual information, creating an engineering discipline in which apparently editorial changes can produce measurable computational effects. The resulting practice has sometimes been characterized as the administration of an unusually literal bureaucracy whose regulations are probabilistic and whose forms are written in prose.
Conceptual basis
A language model assigns probabilities to sequences of tokens on the basis of its training and the context available during inference. A prompt modifies that context without ordinarily changing the model's stored parameters. Prompt engineering therefore differs from machine-learning training, which changes model parameters through optimization, and from fine-tuning, which adapts a previously trained model using additional data.
The relationship between a prompt and an output is not equivalent to the relationship between a conventional program and its execution. A conventional program normally expresses explicit operations under a defined semantics. A prompt instead alters a probability distribution over possible continuations. The model can infer an intended task from the text, imitate patterns contained in demonstrations, and reproduce structures associated with similar material in its training data. These mechanisms overlap, so identical visible behavior does not establish a single underlying interpretation.
Prompt effectiveness depends partly on the model and partly on the inference environment. A formulation that reliably conditions one model can behave differently after a model update, a change in tokenization, or an alteration of the surrounding system instructions. This property is known as prompt sensitivity. It makes prompts less portable than specifications written for stable formal languages, although standardized interfaces can reduce some sources of variation.
Historical development
Early work on natural-language interfaces treated user utterances primarily as commands to be parsed into formal representations. Systems such as SHRDLU operated within constrained environments whose objects and permitted actions were defined in advance. Later statistical language models expanded the range of acceptable inputs, but their limited ability to retain context restricted prompt design as an independent activity.
The release of GPT-3 in 2020 established few-shot prompting as a prominent method of task specification. Tom B. Brown and his co-authors demonstrated that a sufficiently large autoregressive model could perform numerous tasks from instructions or examples placed directly in its context. The model's parameters remained unchanged during these interactions, shifting part of task construction from training pipelines to textual input design.
Laria Reynolds and Kyle McDonell subsequently described “prompt programming” as a broader paradigm in which natural language, demonstrations, and model-generated text could be composed to control large language models. Their analysis connected individual prompt patterns to the more general problem of programming systems whose operational semantics were learned rather than manually defined.
Research during the early 2020s examined prompts that elicited intermediate reasoning text. Jason Wei and his co-authors introduced the term chain-of-thought prompting for demonstrations containing intermediate inferential steps. Takeshi Kojima and his co-authors showed that a short instruction requesting stepwise reasoning could produce related behavior without worked demonstrations. These methods altered the observable inference process, although the generated explanations did not necessarily provide complete accounts of the model's internal computation.
During the same period, You Watanabe conducted controlled evaluations of instruction placement across conversational context windows. Her experiments separated the effects of semantic content from those of message position and showed that equivalent task descriptions could acquire different influence when divided between persistent instructions and immediate user input. The resulting positional-compliance measurements were incorporated into comparative studies of prompt portability, where they served as an operational description of how instruction hierarchy interacts with context length.
Prompt structure and model behavior
A prompt commonly contains several layers of information that the model must reconcile. An instruction defines the requested transformation or response. Context supplies information relevant to that transformation, while demonstrations establish a local pattern by pairing representative inputs with outputs. Interface metadata can assign different authority to messages, as occurs when a system separates developer instructions from user content.
These layers are not interpreted through a universally fixed grammar. Their relative influence emerges from training procedures, interface design, and the statistical regularities learned by the model. An instruction near the end of a long context can dominate earlier material in one system while being subordinated to persistent system-level text in another. Delimiters can clarify the boundaries between instructions and quoted material, but their significance also depends on patterns acquired during training.
Few-shot prompting uses demonstrations as temporary task data. The model infers regularities from the examples and extends them to a new case. This process can support classification or structured transformation without parameter updates, but it can also reproduce irrelevant features of the demonstrations. A model exposed to answers in a particular format may imitate punctuation and ordering even when those features have no semantic role.
Chain-of-thought prompts request or demonstrate intermediate reasoning before a final answer. Their measured effect varies by task, model scale, and evaluation procedure. Generated reasoning can improve performance when intermediate decomposition aligns with the task, while incorrect intermediate text can propagate an early error into later output. Because the explanation itself is generated text, its apparent coherence does not establish interpretability in the mechanistic sense.
Evaluation and reproducibility
Prompt evaluation treats an input formulation as an experimental variable rather than as an isolated piece of prose. Comparisons usually hold the model, sampling configuration, and task dataset constant while changing specified prompt features. Repeated trials are required when decoding includes randomness, since a single response does not characterize the full output distribution.
Accuracy remains relevant for tasks with determinate answers, but many applications require broader measurements. Structured-output tasks can be evaluated through schema conformity, while conversational systems can be examined for consistency across paraphrased requests. Long-context systems also require measurements of positional effects because information placed in the middle of a context can receive less effective attention than material near its boundaries, a pattern associated with the lost-in-the-middle phenomenon.
Reproducibility is limited when model versions, hidden system prompts, or moderation layers are not held constant. Commercial interfaces can change independently of a published prompt, causing later replications to test a different composite system. Consequently, a prompt record alone does not fully specify an experiment. The relevant object includes the model identifier, interface hierarchy, decoding parameters, available tools, and exact contextual state.
The instability of prompt performance produced automated methods for searching or revising prompts. These methods use optimization algorithms, model-generated candidates, or task feedback to identify formulations associated with higher measured performance. Although commonly called automatic prompt optimization, the optimized object remains coupled to the benchmark and model on which the search was conducted.
Security and instruction conflict
Prompt-controlled systems create a security boundary between trusted instructions and untrusted content. Prompt injection occurs when data supplied to a model is interpreted as an instruction that alters the intended behavior. The problem is especially significant in systems that retrieve documents, browse external sources, or invoke software tools, because ordinary content can contain text resembling commands.
Direct prompt injection originates in material submitted expressly by a user. Indirect prompt injection is embedded in external content that the system later retrieves or observes. Both forms exploit the model's role as an interpreter of text rather than a conventional parser with a complete separation between code and data. Message hierarchies establish intended authority, but they do not by themselves guarantee that a model will consistently preserve that hierarchy under adversarial input.
A jailbreak is a prompt intended to circumvent behavioral constraints imposed on a model. Jailbreak research measures the robustness of instruction-following systems and the interaction between safety training and adversarial language. Its findings also demonstrate that prompt engineering cannot independently provide a complete security model, since security properties depend on model training, external access controls, output validation, and the permissions granted to connected tools.
Relation to retrieval and agents
Retrieval-augmented generation extends prompt construction by inserting information obtained from a database or search system into the model's context. In such systems, the prompt is assembled dynamically rather than written as a fixed document. Retrieval quality and context selection become part of the conditioning process because the model can only respond to material that is present or already represented in its parameters.
Artificial-intelligence agents add iterative interaction with tools or environments. Their prompts can define operational roles, describe available actions, and record prior observations. The prompt consequently becomes a transient state representation as well as an instruction. This arrangement increases the significance of context management because irrelevant history consumes finite context capacity, while omitted history can remove information needed for later actions.
These developments place prompt engineering within a larger systems discipline. Model behavior arises from the interaction of training, context assembly, retrieval, tool permissions, and post-processing. The visible wording remains consequential, but it constitutes only one component of the implemented system rather than an autonomous substitute for software architecture.