Every field has its jargon wall, and AI’s is taller than most — tokens, RAG, few-shot, temperature, injection. This prompt engineering glossary knocks that wall down: 36 terms, each explained in plain language, in the order you’ll actually meet them — from your first prompt to advanced territory.
Bookmark this page as your reference while working through the Prompt Engineering roadmap. Terms that deserve a full guide link to one. This is guide #8 — the final piece of the Fundamentals track.
🔤 The Absolute Basics
The text you send to an AI — your question, instruction, or request. The input that everything else is about.
The skill of designing prompts deliberately (words, structure, context, format) so the AI’s output is genuinely useful instead of generic.
The type of AI behind ChatGPT, Claude, and Gemini — a model trained on massive amounts of text to predict what comes next.
The AI’s answer to your prompt. In API contexts you’ll often see it called a ‘completion.’
A specific trained AI (e.g., a particular version of GPT, Claude, or Gemini). Different models have different strengths, speeds, and limits.
🧠 How It Works Under the Hood
The word-chunks AI reads and writes — roughly ¾ of an English word. ‘Cat’ is 1 token; ‘unbelievable’ is 2–3. All limits and pricing count tokens. Full explainer: How AI Prompts Actually Work.
The AI’s working memory — the maximum tokens it can see at once (your prompt + conversation history + its reply). Anything beyond it is forgotten completely.
The act of the model generating a response — ‘running’ the AI on your prompt.
A creativity dial (API setting): low = safe, consistent predictions (facts, code); high = varied, surprising ones (creative work). Chat apps set it for you.
How the model picks the next token from its top candidates — the controlled randomness that makes regenerated answers differ.
A confident but false statement — the model completing a plausible-looking pattern when it lacks the fact. Why you always verify names, numbers, and citations.
The text the model learned patterns from. It explains both what the model knows and why it has a knowledge cutoff.
The date after which the model has no training data — events beyond it are invisible unless the tool has web search.
🛠️ Prompting Techniques
Asking directly with no examples. Fine for common, simple tasks. Guide #9 in our series.
Showing 2–3 examples of the output you want before asking — the strongest quality lever for style and format tasks.
Asking the AI to reason step by step before answering. Dramatically improves math, logic, and decisions.
‘Act as a senior developer / teacher / recruiter’ — sets vocabulary, depth, and standards before the task.
Asking the model to solve a problem multiple ways and pick the most consistent answer — reliability for high-stakes questions.
Having the AI explore several approaches, evaluate them, then choose — for open-ended problems and planning.
Prompts about prompts — e.g., ‘rewrite my prompt to get a better answer.’ The AI improving your input.
Stating what NOT to do: ‘no jargon, no bullet points, don’t invent statistics.’ In image AI, listing what shouldn’t appear.
Breaking a big job into sequenced prompts — outline → draft → edit → format — each step feeding the next.
Treating the first answer as a draft: ‘now shorter,’ ‘more casual,’ ‘add examples.’ The habit that separates pros from beginners.
⚙️ Structure & Settings
Hidden standing instructions that shape an AI’s behavior before your message arrives — the reason the same model can act as different products.
User-level standing preferences some chat apps let you set — a personal mini system prompt applied to every chat.
A reusable prompt with [placeholders] for the parts that change — the building block of a prompt library.
Your saved collection of prompts that worked — the biggest practical productivity unlock in prompting.
The cap on how long a response can be. Hitting it mid-sentence is why answers cut off — say ‘continue.’
Models that accept more than text — images, audio, files — as part of the prompt.
🚀 Advanced Territory
Giving the model relevant documents at question-time so it answers from YOUR data instead of memory — the backbone of ‘chat with your PDF’ tools. Deep dive: RAG Prompting guide.
A security attack where malicious text inside content tricks the AI into ignoring its instructions — the reason AI apps must treat user content carefully.
Attempts to trick a model into bypassing its safety rules. Understanding the term matters for anyone building AI products.
Retraining a model on custom examples to change its default behavior — compare with prompting, which changes behavior per-request. See: Fine-Tuning vs Prompt Engineering.
An AI that doesn’t just answer but takes multi-step actions — searching, using tools, executing plans. Prompting agents is its own skill.
Numbers that represent meaning — how machines measure that ‘car’ and ‘automobile’ are similar. Powers search and RAG.
Anchoring the model’s answer in provided sources (documents, search results) rather than memory — the antidote to hallucination.
Fundamentals track: complete 🎓
Next stop — the Core Techniques track on the roadmap.
Frequently Asked Questions
Five carry you a long way: prompt, token, context window, few-shot, and iteration. Understand those and every guide in this field becomes readable.
A token is a chunk of text the model processes — often a whole short word, but longer words split into pieces. On average, 100 tokens ≈ 75 English words.
Retrieval-Augmented Generation — the AI is handed relevant documents at question time and answers from them instead of memory. It’s how ‘chat with your PDF’ tools work.
Related but different: jailbreaking is a user trying to bypass the model’s rules directly; prompt injection hides malicious instructions inside content (a webpage, an email) that the AI processes.
No — daily prompting needs only the basics and techniques sections. The advanced terms matter when you start building AI tools or reading technical discussions.
