Prompt Engineering Glossary

A

Accuracy check #

A review step that compares a model's answer with a trusted source to confirm that facts are correct.

Agent #

An AI system that plans steps, uses tools and takes actions toward a goal, usually guided by prompts and feedback.

Answer format control #

Specifying how the answer should look, such as a summary, numbered steps or a comparison table.

Aspect ratio parameter #

A setting in image or video prompts that fixes the frame shape, such as 16:9 or 9:16.

Attention mechanism #

The part of a transformer model that weighs how much each token should influence the others when producing output.

Audience prompt #

An instruction that tells the model who the answer is for, which changes vocabulary and depth.

Audio prompt #

A description of sound, music or ambience requested from a generative audio model.

B

Batch processing #

Sending many prompts together for processing at a lower cost, usually with results returned later.

Beam search #

A decoding method that keeps several candidate continuations at each step and selects the most likely complete sequence.

Benchmark #

A standard set of tasks and scoring rules used to compare how well different models or prompts perform.

Brand voice prompt #

An instruction that defines a brand's tone, vocabulary and style so generated text stays consistent.

Brief template #

A structured prompt that collects goal, audience, tone, constraints and format for a task.

C

Camera motion prompt #

A prompt instruction that specifies a camera move, such as a slow dolly in or an upward crane.

Chain of verification #

A technique in which the model drafts an answer, writes checks against it and then revises the answer.

Chain-of-thought prompting #

A technique that asks the model to show intermediate reasoning steps before giving a final answer.

Character consistency #

Keeping a character's face, clothing and proportions the same across multiple generated images or clips.

Character reference #

An image supplied with a prompt to keep a character's appearance consistent.

Chunking #

Splitting a long document into smaller sections so each can be retrieved or processed within a model's limits.

Clarifying question #

A question the model asks when a prompt is ambiguous, before it produces an answer.

Classifier-free guidance #

A setting in diffusion image and video models that controls how strongly the output follows the prompt.

Concise output #

A prompt constraint that asks for brevity and removes unnecessary detail.

Constraint #

A rule in a prompt that limits the output, such as length, format, banned words or required sections.

Context stuffing #

Placing large amounts of material into a prompt, which can dilute attention and raise cost.

Context window #

The maximum amount of text, measured in tokens, that a model can consider in one request.

D

Decoding #

The process by which a model selects output tokens one by one from its probability distribution.

Decomposition #

Splitting a complex task into smaller sub-tasks that are each easier to prompt and check.

Delimited input #

User or source text placed inside clear markers so the model can tell instructions from content.

Delimiter #

A marker such as triple quotes, XML tags or dashes that separates instructions from input text in a prompt.

Diffusion model #

A generative model that creates images or video by removing noise step by step from random input.

Digital watermark #

A hidden signal added to AI-generated media to show that it was created or altered by AI.

Distillation #

Training a smaller model to imitate the outputs of a larger model, keeping most of its ability at lower cost.

E

Edge case #

An unusual input that tests whether a prompt still behaves correctly.

Embedding #

A numerical vector that represents the meaning of a word, sentence or document, used for search and comparison.

Error analysis #

Reviewing failed outputs to find patterns and adjust the prompt or data.

Evaluation harness #

A set of test prompts, expected results and scoring code used to measure prompt or model performance.

Example selection #

Choosing the most representative examples for a few-shot prompt.

Exemplar #

An example input and output included in a prompt to show the desired behavior.

Expected output #

The correct or acceptable response against which a prompt's result is compared.

F

Factuality #

The degree to which a model's statements match verifiable facts.

Few-shot examples count #

The number of examples included in a few-shot prompt, which affects cost and accuracy.

Few-shot prompting #

Including a small number of examples in the prompt so the model learns the pattern before answering.

Fine-tuning #

Further training a pre-trained model on a specific dataset to change its behavior or style.

Format instruction #

A line in a prompt that defines how the output must be arranged, such as headings or bullet points.

Frame interpolation prompt #

A request to generate intermediate frames to smooth motion in a generated video.

Function calling #

A feature that lets a model request a structured call to an external tool, such as a database query or API.

G

Golden dataset #

A trusted set of inputs with verified correct outputs, used to test prompts and models over time.

Groundedness #

The degree to which each claim in a response is supported by the supplied sources.

Grounding #

Connecting a model's answers to supplied sources or data so that responses are based on verifiable information.

Guardrail #

A rule, filter or check that prevents unsafe, off-topic or non-compliant outputs.

H

Hallucination #

An output that states something false or unsupported as if it were true.

Hallucination rate #

The share of outputs that contain unsupported or false statements in a test set.

Human-in-the-loop review #

A process in which a person checks and approves model outputs before they are used.

I

Image-to-video prompt #

A prompt that describes how a still image should move when turned into a short video.

In-context learning #

A model's ability to adapt its behavior from examples or instructions provided within the prompt.

Instruction hierarchy #

The order in which a model should prioritise system, developer and user instructions when they conflict.

Instruction tuning #

Training a model on examples of instructions and good responses so it follows prompts more reliably.

J

Jailbreak #

An attempt to make a model ignore its safety rules through crafted prompts.

JSON mode #

A setting that constrains a model to return valid JSON, making its output easier for software to parse.

K

Knowledge base #

A collection of approved documents that a retrieval system searches to ground model answers.

Knowledge cutoff #

The date after which a model has no training data, so it may not know about newer events.

L

Language constraint #

A rule that sets the language, reading level or terminology the output must use.

Latency #

The time between sending a prompt and receiving the first or complete response.

Length constraint #

A rule that limits the number of words, sentences or paragraphs in a response.

Lens prompt #

A description of focal length or lens type, such as a 35mm lens, used to control framing and perspective.

Lighting prompt #

A description of light direction, quality and color used to control the look of a generated image or video.

LLM-as-judge #

A method that uses one language model to score or compare the outputs of another against a rubric.

Logit #

The raw score a model assigns to each possible next token before it is converted into probabilities.

M

Max tokens #

A setting that limits the length of a model's response, measured in tokens.

Meta prompt #

A prompt that asks a model to write or improve another prompt.

Model card #

A document describing a model's intended uses, limits, training data and known risks.

Motion prompt #

A description of movement for a subject or object in a generated video.

Multimodal prompt #

A prompt that combines text with other inputs such as images, audio or video.

N

Negative example #

An example of an unwanted output included in a prompt to show what to avoid.

Negative prompt #

A description of what the output should avoid, used in image and video generation.

Non-determinism #

The tendency of a model to produce different outputs for the same prompt, due to sampling.

O

Output format #

The structure a response must follow, such as a table, bullet list, JSON object or specific headings.

Output schema #

A formal definition of the fields, types and required items in a structured response.

Output token #

A token generated by the model in its response, which is usually billed separately from input tokens.

P

Perplexity #

A measure of how well a language model predicts a sample of text; lower values indicate better prediction.

Persona consistency #

The ability of a model to keep the same tone, knowledge and role throughout a conversation.

Persona prompting #

Asking the model to respond as a specific role, such as a senior editor or a customer service agent.

Pilot test #

A small trial of a prompt on real inputs before it is rolled out widely.

Pipeline #

A sequence of prompts, tools and checks that together complete a task.

Plan-and-execute #

A pattern in which the model first writes a plan and then carries out each step.

Politeness token #

A polite phrase in a prompt that has little effect on output quality compared with clear instructions.

Positional bias #

A tendency for a model to give more attention to information at the start or end of a long prompt.

Prompt #

The input text, instructions or media given to a model to produce an output.

Prompt audit #

A review of prompts for clarity, safety, cost and consistency with brand and policy.

Prompt caching #

Reusing a stored, unchanged part of a prompt so that repeated requests cost less and respond faster.

Prompt chaining #

Breaking a task into a series of prompts where each output becomes the input for the next step.

Prompt compression #

Shortening a prompt while keeping its essential meaning, to reduce cost and latency.

Prompt debugging #

Finding and fixing the part of a prompt that causes an incorrect or unexpected output.

Prompt drift #

A decline in prompt performance over time, often caused by model updates or changing inputs.

Prompt evaluation #

Measuring how well a prompt performs against defined criteria and test cases.

Prompt injection #

An attack in which hidden instructions in user input or retrieved content override the original prompt.

Prompt leakage #

The unintended disclosure of a system prompt or hidden instructions through a model's output.

Prompt library #

A stored, organized collection of tested prompts that a team can reuse and improve.

Prompt quality score #

A rating that combines clarity, accuracy and format compliance to compare prompts.

Prompt template #

A reusable prompt with placeholders that are filled in with specific data for each request.

Prompt testing #

Running a prompt against a set of inputs to check that its outputs meet the requirements.

Prompt versioning #

Recording each change to a prompt with a version number so results can be compared and rolled back.

Prompt weighting #

Assigning greater or lesser emphasis to words or phrases in a prompt, supported by some image and video tools.

R

Rate limit #

A cap on the number of requests or tokens that can be sent to a model within a time period.

Reasoning model #

A model designed to spend extra computation on step-by-step reasoning before it answers.

Redaction #

Removing personal or confidential information from a prompt or output before it is stored or shared.

Reference image #

An image given to a model to guide style, composition or subject in a generated output.

Reflection #

A technique in which the model reviews and critiques its own output before producing a final answer.

Regression test #

A repeated test that checks whether a prompt still works after a change to the prompt or the model.

Retrieval query #

A search request, often generated from the user's question, that finds relevant passages for a model.

Retrieval-augmented generation #

A method that retrieves relevant documents and adds them to the prompt before the model answers.

Role prompting #

Assigning the model a role, such as analyst or teacher, to shape its tone and expertise.

Rubric #

A set of criteria and scoring levels used to judge the quality of a response.

S

Safety filter #

An automated check that blocks or flags content considered harmful before or after generation.

Sampling #

Choosing the next token at random according to its probability, which introduces variation in outputs.

Scratchpad #

A space in the prompt or output where the model writes intermediate work before its final answer.

Seed #

A number that initialises random generation so that a result can be reproduced or varied in a controlled way.

Self-consistency #

Generating several reasoning paths for the same question and selecting the most common answer.

Short-form output #

A response limited to a few sentences or a single short item, such as a headline or caption.

Shot prompt #

A prompt written for a single shot in a sequence, usually describing subject, framing, action and setting.

Source attribution #

Naming the documents or references that support a model's claims.

Stakeholder review #

A review of AI outputs by the people who will use or approve them, such as clients or editors.

Stop sequence #

A string that tells the model to end its output when it appears.

Storyboard prompt #

A prompt that describes a sequence of frames to generate a planned set of shots.

Structured output #

A response returned in a defined format, such as JSON or a table, so it can be processed by software.

Style guide prompt #

A prompt that includes a brand's writing or visual rules so outputs follow them consistently.

Style reference #

An image or description supplied to guide the visual style of a generated output.

System prompt #

Instructions that set a model's overall behavior, role and rules for a conversation or application.

T

Task instruction #

The part of a prompt that tells the model exactly what to do.

Temperature #

A setting that controls how random a model's output is; lower values give more predictable results.

Temperature zero #

A setting that makes output as predictable as possible by always choosing the most likely token.

Test case #

A single input with an expected result, used to check a prompt.

Text-to-image prompt #

A prompt that describes an image to be generated from text alone.

Text-to-video prompt #

A prompt that describes a video to be generated, including subject, action, setting, camera and style.

Token #

A unit of text, often a word fragment, that a model reads and generates; usage and cost are measured in tokens.

Token budget #

The limit on the number of tokens allowed for a prompt and response, used to manage cost and length.

Token efficiency #

Achieving the same result with fewer tokens, which lowers cost and latency.

Tone instruction #

A line in a prompt that defines the emotional tone and formality of the response.

Tool use #

A model's ability to call external functions, search engines or software to complete a task.

Top-k sampling #

A method that restricts random choice to the k most likely next tokens.

Top-p sampling #

A method that restricts random choice to the smallest set of tokens whose probabilities add up to p.

Transparency note #

A statement telling readers that content was generated or assisted by AI.

U

Unstructured input #

Free text, such as notes or emails, supplied to a model without fixed fields.

V

Validation prompt #

A prompt that asks a model to check another output against rules or a schema.

Vector database #

A database that stores embeddings and finds the items most similar in meaning to a query.

Version pinning #

Locking a project to a specific model version so outputs do not change when the provider updates the model.

Video consistency #

The degree to which subjects, lighting and style remain stable across frames and shots.

Video length prompt #

An instruction that states the intended duration of a generated clip.

Voice prompt #

A description of tone, pace, accent and emotion for a generated or synthetic voice.

Z

Zero-shot chain-of-thought #

Asking the model to reason step by step without providing examples, using a simple instruction such as think step by step.

Zero-shot prompting #

Asking a model to perform a task without providing any examples in the prompt.