Prompt engineering has its own vocabulary, and half the confusion marketers run into comes from terms used loosely or inconsistently across blog posts and tool documentation. This glossary defines the terms that actually matter for marketing work, grouped by category, in plain language rather than academic phrasing.
Treat this as a working reference: skim the category you need, not the whole thing front to back.
Core Prompt Engineering Terms
Prompt — the instruction or input you give a language model to produce a response. Everything from a single sentence to a multi-paragraph brief counts as a prompt.
Prompt engineering — the practice of deliberately designing, testing, and refining prompts to reliably get useful, on-brand output rather than generic or inconsistent results.
System prompt — a standing instruction set applied to every interaction in a session, often used to establish persistent rules like tone, banned phrases, or output format, separate from the specific task request.
Context window — the total amount of text, measured in tokens, a model can consider at once, including your prompt, any reference material, and its own response. Exceeding it means older content gets dropped or ignored.
Token — a chunk of text, roughly three-quarters of a word on average, that models use as their basic unit for reading input and generating output. Prompt length and cost are usually measured in tokens.
Instruction following — how reliably a model adheres to the specific constraints in a prompt, such as word counts or formatting rules, rather than drifting toward its own default habits. Weaker instruction following is one of the main reasons the same prompt can produce inconsistent results across attempts.
Completion — the model’s generated response to a given prompt, sometimes still called this even in conversational tools, a holdover term from earlier text-completion-style models.
Prompt Structure and Formatting Terms
Zero-shot prompting — asking the model to complete a task with no examples provided, relying entirely on the instruction itself.
Few-shot prompting — including a small number of examples of the desired output directly in the prompt, which typically improves consistency and tone-matching over zero-shot prompting.
Chain-of-thought prompting — instructing a model to reason through a problem step by step before producing a final answer, useful for tasks involving logic, comparison, or multi-part analysis rather than straightforward drafting.
Role prompting — assigning the model a specific persona or expertise framing (such as “act as a local SEO strategist”) to shape the angle and vocabulary of its response.
Delimiter — a marker used inside a prompt to clearly separate instructions from reference material or input data, reducing the chance the model confuses the two.
Model Behavior and Output Terms
Hallucination — when a model generates information that sounds plausible but is factually wrong or entirely fabricated, such as an invented statistic or a nonexistent source. This is the single biggest risk in AI-assisted marketing content and the main reason human fact-checking remains non-negotiable.
Temperature — a setting that controls how predictable versus varied a model’s output is. Lower temperature produces safer, more consistent responses; higher temperature produces more varied, sometimes more creative, but less predictable results.
Grounding — providing a model with specific, verifiable source material to base its response on, reducing the likelihood of hallucination compared to relying on the model’s general training knowledge alone.
Output drift — the tendency for a model’s responses to gradually lose adherence to earlier instructions over a long conversation, a common reason to periodically restate key rules in longer sessions.
Advanced Prompting Techniques
Prompt chaining — breaking a complex task into a sequence of smaller prompts, where the output of one becomes the input to the next, rather than attempting everything in a single instruction.
Iterative refinement — the practice of testing a prompt, evaluating the output against a specific standard, and adjusting the prompt based on identified gaps, repeated until output quality is reliably strong.
Negative prompting — explicitly instructing a model what to avoid (banned phrases, structures, or claims) rather than only specifying what to include, often the fastest fix for recurring output problems.
Constraint-based prompting — building hard limits directly into a prompt, such as word count, character count, or required structural elements, to reduce the need for post-generation editing.
AI Search and SEO-Specific Terms
AI Overview — a search engine’s AI-generated summary shown above traditional organic results, synthesizing information from multiple sources rather than linking to a single page.
Answer engine optimization — the practice of structuring content so it’s easily extracted and cited by AI-generated answers and summaries, related to but distinct from traditional ranking-focused SEO.
Citation — in an AI-search context, a source referenced or linked within an AI-generated answer, functionally similar to earning a featured snippet in traditional search.
Extractability — how easily a piece of content can be pulled out and summarized by an AI system, typically improved by clear structure, direct answers near the top of a section, and unambiguous phrasing.
Workflow and Quality Control Terms
Prompt library — a stored, reusable collection of tested prompts organized by task, allowing a team to produce consistent output without rebuilding instructions from scratch each time.
Human-in-the-loop — a workflow where a person reviews, edits, or approves AI-generated output before it’s used or published, standard practice for any client-facing or published marketing content.
Voice brief — a document translating a brand’s tone and style preferences into concrete, model-readable instructions, more useful than vague adjectives like “friendly” or “professional.”
Anti-pattern — a recurring prompting mistake, such as vague instructions or missing context, that reliably produces weak or generic output and should be flagged and corrected in a prompt library.
Edit distance — a practical measure of how much a draft changes between AI generation and final publication, used as a proxy for how well a prompt is actually performing rather than how fast it produced a first draft.
Version control (for prompts) — the practice of tracking changes to a prompt over time, similar to code versioning, so a team can see what changed, when, and why a given revision improved or hurt output quality.
Terms You'll Hear but Rarely Need
Fine-tuning — retraining a model on a custom dataset so its default behavior changes permanently, distinct from prompt engineering, which shapes output at the moment of use without altering the underlying model. Most marketing teams never need this; a well-built prompt handles the vast majority of use cases.
Embedding — a numerical representation of text used by systems to compare meaning and relevance, relevant mostly to search and retrieval tooling rather than day-to-day prompting itself.
Retrieval-augmented generation — a technique where a system automatically pulls relevant reference material into a prompt before generating a response, often used under the hood by tools that answer questions from a specific knowledge base rather than general training data.
Frequently Asked Questions
What's the difference between a prompt and prompt engineering?
A prompt is a single instruction given to a model. Prompt engineering is the ongoing discipline of designing, testing, and refining prompts to reliably produce quality, on-brand output rather than relying on one-off guesses.
Is few-shot prompting always better than zero-shot?
Not always — zero-shot is faster and often sufficient for simple, well-understood tasks. Few-shot prompting earns its added complexity when tone-matching or format consistency really matters, such as brand voice work.
What causes hallucinations, and can prompt engineering prevent them?
Hallucinations happen when a model generates plausible-sounding but unverified information, often when a prompt lacks grounding in real source material. Strong prompting reduces the risk significantly but never eliminates the need for human fact-checking.
Do I need to understand tokens and context windows to write good prompts?
A basic understanding helps, particularly for longer tasks involving large reference documents, but most marketers can prompt effectively without deep technical knowledge of how tokens are counted.
What's the fastest way to fix a prompt that keeps producing generic output?
Add specific, real reference material and explicit negative prompting — telling the model what to avoid — which usually resolves genericness faster than simply asking for "more creative" or "better" output.
Terry has 30+ years in software and SEO. He’s the founder of Salterra Digital Services and SEO Spring Training, host of the Roundtable SEO Mastermind, and lead instructor at SEO University — teaching the exact tactics his team uses on client work.
Continue learning
Ready to master this?
This guide is one lesson from the Prompt Engineering course. Get every lesson, framework and checklist — plus the full 38-course catalog — inside SEO University.
Practitioner-focused training across the full digital marketing stack — from technical SEO to conversion optimization and the AI search era. By Salterra Digital Services, since 2011.