AI Prompt Engineering Glossary: 40+ Terms Explained (2026)

**Prompt.** The full input text you send a model — instructions, context, and any data — that conditions its output.

**Prompt engineering.** The practice of designing and refining prompts to get reliable, useful outputs. See the DAIR.ai guide.

**Zero-shot prompting.** Asking the model to do a task with no worked examples, relying only on instructions and its pretraining.

**Few-shot prompting.** Including one or more input/output examples in the prompt to steer format and behavior. Popularized by Brown et al. 2020 (arXiv:2005.14165).

**One-shot prompting.** Few-shot with exactly one example.

**In-context learning.** A model adapting to a task from examples shown in the prompt, without any weight updates — the mechanism behind few-shot prompting.

**Chain-of-Thought (CoT).** Prompting the model to reason through intermediate steps before answering, which improves performance on multi-step problems. Introduced by Wei et al. 2022 (arXiv:2201.11903).

**Zero-shot CoT.** Eliciting step-by-step reasoning without examples, often by adding a cue like 'Let's think step by step.'

**Self-consistency.** Sampling several reasoning paths and taking the most common answer, rather than trusting a single chain.

**Self-critique / reflection.** Having the model review and revise its own output to catch errors before finalizing.

**ReAct.** A pattern that interleaves reasoning ('Thought') with actions like tool calls ('Action'), looping until the task is done. From Yao et al. 2023 (arXiv:2210.03629).

**Tree of Thoughts (ToT).** Exploring multiple reasoning branches and evaluating them, rather than committing to one linear chain. From Yao et al. 2023 (arXiv:2305.10601).

**Prompt chaining.** Breaking a task into a sequence of prompts where each step's output feeds the next.

**Tool use / function calling.** Giving a model a defined set of tools (functions, APIs) it can choose to invoke, with the results fed back into context.

**Agent.** A system that uses an LLM to plan and take actions toward a goal, typically with tools, memory, and an iteration loop.

**Plan-and-execute.** An agent pattern that first generates a plan, then executes the steps — often with a cheaper model running the steps.

**System prompt.** A high-priority instruction block that sets a model's role, rules, and behavior for a conversation, separate from the user's turns.

**Role prompting.** Assigning the model a persona or job ('You are a senior editor') to constrain tone and expertise.

**Meta-prompting.** Using a model to help write or improve prompts themselves.

**RAG (Retrieval-Augmented Generation).** Fetching relevant documents at query time and inserting them into the prompt so the model answers from supplied evidence rather than memory alone.

**Grounding.** Tying a model's output to specific source material so claims can be traced and verified.

**Embedding.** A numeric vector representing the meaning of text, used to find semantically similar content.

**Vector search / semantic search.** Retrieving content by similarity of embeddings rather than exact keyword match.

**Chunking.** Splitting source documents into smaller passages so the most relevant pieces can be retrieved and fit in context.

**Hallucination.** When a model produces confident but false or unsupported content. Grounding and RAG reduce it.

**Citation.** A source reference attached to a model's claim so a reader can verify it — central to search-grounded tools.

**Token.** The unit a model reads and generates — roughly 4 characters or about 0.75 words of English, per Anthropic and OpenAI docs (a rough estimate).

**Tokenization.** The process of splitting text into tokens before the model processes it.

**Context window.** The maximum number of tokens (prompt plus output) a model can consider at once.

**Temperature.** A sampling setting that controls randomness — lower is more deterministic, higher is more varied. See the OpenAI API reference.

**Top_p (nucleus sampling).** Restricts sampling to the smallest set of tokens whose probabilities sum to p, an alternative way to control diversity. See the OpenAI API reference.

**Top_k.** Restricts sampling to the k most likely next tokens.

**Max tokens.** A cap on how many tokens the model may generate in a response.

**Stop sequence.** A string that, when generated, tells the model to stop producing output.

**Logit bias.** A per-token adjustment that makes specific tokens more or less likely.

**Greedy decoding.** Always choosing the single most probable next token (equivalent to temperature 0).

**Output format / structured output.** Instructing the model to return a specific shape, such as JSON with named keys or a markdown table.

**Delimiters.** Markers (triple backticks, XML-style tags, headings) that separate instructions from data within a prompt.

**Constraint.** An explicit rule on the output — length, tone, things to include or avoid.

**Prompt template.** A reusable prompt with placeholders you fill in per task.

**Negative prompting.** Telling the model what to exclude or avoid; common in image generation. See the DAIR.ai guide.

**Fallback instruction.** A rule for missing or ambiguous inputs (e.g. 'if a value isn't present, write null') that prevents fabricated values.

**Prompt caching.** Reusing the processed form of a stable prompt prefix to cut cost and latency on repeated calls; cache reads are typically a fraction of normal input price.

**Batch processing.** Submitting many requests together for a discount, often at the cost of latency.

**Fine-tuning.** Further training a base model on task-specific data to change its behavior, distinct from prompting.

**Evaluation (evals).** Systematic testing of model or prompt quality against a fixed set of cases.

**Guardrails.** Rules and filters that constrain model inputs or outputs for safety and policy compliance.

**Prompt injection.** An attack where malicious instructions hidden in input data hijack the model's behavior. Ranked the #1 risk (LLM01:2025) on the OWASP LLM Top 10.

**System prompt leakage.** Exposure of a system prompt's contents to users or attackers — ranked LLM07:2025 on the OWASP LLM Top 10.

**Jailbreak.** A prompt crafted to bypass a model's safety guidelines or guardrails.