AI Prompt Vocabulary 101: 20 Terms Every Beginner Needs Before Their First Prompt

The text you send to an AI model to tell it what to do. It can be a question, an instruction, a document plus a question about that document, or a whole set of rules. Merriam-Webster and Oxford both track “prompt” as a fast-moving word of the 2020s; it was runner-up for Oxford’s word of the year in 2023.

Example: “Summarize this email in two sentences for my manager.” — that whole sentence is the prompt.

A separate instruction, usually invisible to the end user, that sets the model’s role, tone, and rules before the conversation starts. When a chatbot always answers in a certain persona or refuses certain topics no matter what you type, that behavior almost always comes from its system prompt, not from you.

Example: “You are a customer support agent for a bakery. Never discuss competitors. Keep replies under 80 words.”

The unit a model actually reads and writes in — a chunk of text that’s often shorter than a whole word (part of a word, a whole word, or a punctuation mark). A widely cited rule of thumb is that one token is roughly four characters, or about three-quarters of an English word. Everything you type, and everything the model replies with, gets billed and measured in tokens, not characters or words.

In practice: a 1,000-word blog post is roughly 1,300 tokens.

The total number of tokens a model can “see” at once — your prompt, any files you attach, the conversation so far, and its own reply, all counted together. Once a conversation exceeds the window, the oldest content typically gets dropped or summarized, and the model genuinely can’t see it anymore. This is a hardware-and-training limit, not a setting you toggle.

Watch for: a chatbot that “forgets” something you said forty messages ago probably didn’t get confused — it ran out of window.

Asking a model to do a task with no examples at all — you describe what you want and trust the model’s training to fill in the rest. Most casual chatbot use is zero-shot by default.

Example: “Translate this sentence into formal French.” (No sample translation given.)

Showing the model two or three examples of the input-output pattern you want before giving it the real task. This technique was formally described in the 2020 GPT-3 research paper “Language Models are Few-Shot Learners,” and it remains one of the most reliable ways to control output format.

Example: Show two sample product titles rewritten in your brand voice, then ask the model to rewrite a third in the same pattern.

Asking the model to reason step by step before giving its final answer, rather than jumping straight to a conclusion. It measurably helps with multi-step math, logic, and planning tasks, and it’s one of the most-cited techniques in the academic prompting literature — a 2024 systematic survey catalogued over 50 distinct text-based prompting techniques, and chain-of-thought variants are among the best studied.

Example: “Work through this step by step, then give your final answer on the last line.”

Telling the model to respond as a particular role — “act as an editor,” “act as a skeptical reviewer.” It’s genuinely useful for setting tone and scope. One honest caveat worth knowing before you rely on it: several 2026 practitioner write-ups on production AI systems report that elaborate “you are a world-class expert” persona preambles show little to no measurable accuracy benefit on current frontier models, and mainly cost you tokens. Use role prompting for tone and framing, not as a magic accuracy switch.

Example: “Act as a copy editor. Only flag grammar and clarity issues, not style preferences.”

A number, typically between 0 and 1 (sometimes higher), that controls how predictable the model’s word choices are. Low temperature makes the model pick its single most likely next word almost every time — consistent, sometimes repetitive. High temperature lets it choose less-likely words more often — more varied, occasionally less coherent.

Rule of thumb: low temperature (0–0.3) for factual or code tasks, higher (0.7–1) for brainstorming and creative writing.

A second, related dial. Instead of a temperature “heat” setting, top-p limits the model to only the smallest set of next-word candidates whose combined probability adds up to p. A top-p of 0.9 means the model only considers words from the most likely 90% of the probability mass, ignoring the long tail of unlikely options. Most guides recommend adjusting temperature or top-p, not both at once, since they interact.

A hard ceiling on how many tokens the model is allowed to generate in its reply. Set it too low and you’ll get an answer cut off mid-sentence; set it needlessly high and you may pay for or wait on a longer response than you need.

A specific string of text that tells the model to stop generating the moment it produces that string. Developers use this to keep a model from rambling past a clean structural boundary, like the end of a JSON object or a specific delimiter.

Example: setting “\n\n” as a stop sequence to end generation after one paragraph.

When a model states something false, fabricated, or unverifiable with the same confident tone it uses for true statements — a made-up citation, a wrong date, a plausible-sounding but nonexistent product feature. It happens because the model is predicting statistically likely text, not looking facts up in a database, unless it’s specifically been given a tool to do that.

Mitigation: ask the model to say “I’m not certain about this” instead of guessing, or connect it to a real source via retrieval (see term 18, below).

An attack where hidden or malicious instructions are smuggled into content the model reads — a webpage, a document, an email — in an attempt to override the model’s actual instructions. It’s a live, actively studied security concern for any AI system that reads untrusted content, not a theoretical risk.

A prompt specifically crafted to get a model to bypass its safety training and produce content it’s designed to refuse. Related to, but distinct from, prompt injection: a jailbreak is usually typed directly by the user; an injection usually arrives through content the model is reading on the user’s behalf.

A 2025–2026 term for a specific, counterintuitive failure: even models with very large context windows produce measurably worse answers once you fill that window with a lot of marginally relevant content, well before you hit the technical limit. Practitioners in 2026 increasingly describe the fix as “curate, don’t dump” — feeding a model less but more relevant material outperforms feeding it everything you have.

A system where the model doesn’t just answer once, but plans, takes actions using tools, checks the results, and adjusts — looping until a task is done rather than stopping after one reply. “Research these competitors, build a comparison table, then draft a positioning memo” is an agentic task; a single Q&A exchange is not.

A technique that gives a model access to an external, up-to-date knowledge source — a document set, a database, a live search — and has it pull relevant material into its context before answering, instead of relying purely on what it memorized during training. It’s the single most common way production systems reduce hallucination on factual and company-specific questions.

Rules, filters, or checks — inside the prompt, or wrapped around the model as separate code — that keep an AI system from producing harmful, off-topic, or policy-violating output. A refusal you see from a chatbot is often a guardrail firing, not the base model “deciding” on its own.

The discipline covered in the section above: deliberately designing what information — documents, memory, tool outputs, conversation history — lands in a model’s context window at each step, instead of just refining the wording of a single instruction. If you remember one thing after finishing this glossary, it’s that this term now sits one layer above everything else on this page.