Prompt Engineering Glossary
Master the terminology of generative AI. From Attention Mechanisms to Zero-Shot Prompting, explore 100+ essential concepts.
A
Crafting prompts designed to test or bypass AI safety guardrails.
Security researchers use adversarial prompts to identify vulnerabilities in LLM outputs.
The neural network component that allows models to focus on relevant parts of input text.
Transformer attention helps the model understand context across long prompts.
B
A model where users provide their own API keys for AI services, maintaining privacy and cost control.
AI Prompt Architect uses BYOK so your API keys never touch our servers.
A decoding strategy that explores multiple output sequences simultaneously to find optimal responses.
Beam search with width 5 considers the top 5 token sequences at each step.
C
A prompting technique that asks the model to show its reasoning step by step before giving a final answer.
Adding 'Let's think step by step' improved math accuracy from 58% to 93%.
The maximum number of tokens an AI model can process in a single interaction.
GPT-4o has a 128K context window; Claude 3.5 supports 200K tokens.
The text generated by an AI model in response to a prompt.
The completion for 'Write a haiku about coding' would be the generated haiku.
D
Breaking a complex task into smaller, manageable sub-prompts for better results.
Instead of 'write a business plan', decompose into: executive summary, market analysis, financial projections.
Special characters used to clearly separate different sections within a prompt.
Using triple backticks (```) to separate code from instructions.
E
A numerical vector representation of text that captures semantic meaning.
Similar concepts like 'dog' and 'puppy' have embeddings close together in vector space.
F
Providing several examples in the prompt to guide the model's output format and style.
Showing 3 example customer reviews with sentiment labels before asking the model to classify new ones.
Training a pre-trained model on a specific dataset to specialize it for particular tasks.
Fine-tuning GPT-3.5 on legal documents to create a specialised legal assistant.
G
Connecting AI responses to factual, verifiable information sources.
Using RAG to ground responses in company documentation reduces hallucinations.
Safety mechanisms that prevent AI models from generating harmful or inappropriate content.
System prompts can act as guardrails: 'Never provide medical diagnoses'.
H
When an AI model generates plausible-sounding but factually incorrect information.
The model confidently cited a research paper that doesn't exist.
I
The ability of LLMs to learn new tasks from examples provided within the prompt, without weight updates.
Showing the model 5 translation pairs lets it learn your preferred translation style.
Training a model to follow natural language instructions more effectively.
InstructGPT was trained to be better at following user instructions than base GPT-3.
J
Attempting to bypass an AI model's safety restrictions through creative prompting.
Researchers test jailbreaks to improve model safety, not for malicious use.
L
A neural network trained on massive text datasets, capable of understanding and generating human language.
GPT-4, Claude, and Gemini are all large language models.
An efficient fine-tuning method that trains only a small number of additional parameters.
LoRA lets you fine-tune Llama on a single GPU instead of requiring a cluster.
M
AI models that can process and generate multiple types of data (text, images, audio, video).
GPT-4o is multi-modal — it can analyse images and respond with text.
A parameter that limits the maximum length of the model's response.
Setting max_tokens=500 ensures responses stay concise.
P
The input text or instructions given to an AI model to generate a desired response.
A well-crafted prompt includes context, task, constraints, and output format.
Connecting multiple prompts in sequence, where each output feeds into the next prompt.
First prompt extracts data, second analyses it, third generates a report.
R
A technique that retrieves relevant documents before generating a response, reducing hallucinations.
RAG systems search your knowledge base first, then use those results to inform the AI's answer.
A training technique where human preferences guide the model towards more helpful responses.
ChatGPT was improved using RLHF — human raters ranked outputs to train a reward model.
Assigning a specific persona or expertise to the AI to improve response quality.
'You are a senior tax accountant with 20 years of UK experience.'
S
A four-part prompt structure: Situation, Task, Constraints, Output — designed for consistent, high-quality AI interactions.
STCO users report 73% better results than freeform prompting.
A special prompt that sets the AI's behaviour, personality, and rules for an entire conversation.
System: 'You are a helpful coding assistant. Always provide code examples.'
A technique that generates multiple responses and selects the most common answer for higher accuracy.
Asking the model to solve a math problem 5 times and taking the majority answer.
T
A parameter controlling randomness in AI outputs. Lower = more focused, higher = more creative.
Temperature 0.1 for factual tasks, 0.8 for creative writing.
The basic unit of text that AI models process — roughly 4 characters or ¾ of a word.
'Hello world' is 2 tokens. 'Pneumonoultramicroscopicsilicovolcanoconiosis' is 9 tokens.
A decoding method that considers only the K most probable next tokens.
Top-K=40 means the model picks from its 40 best guesses at each step.
A decoding method that considers tokens until the cumulative probability reaches P.
Top-P=0.9 considers the smallest set of tokens whose probabilities sum to 90%.
Z
Master these concepts in practice
Stop reading and start building. Try our STCO framework editor today.
We value your privacy
We use cookies and similar technologies to ensure our website works properly, analyze traffic, and personalize your experience. Under the GDPR, CCPA, and CPRA, you have the right to choose which categories, apart from necessary cookies, you allow.