Prompt Engineering flashcards, generated for you.

You're having a conversation with an AI to get it to understand what you want. Better instructions → better answers, just like explaining something clearly to a person.

Language models predict the next token based on patterns in training data. Clearer prompts activate the right patterns; ambiguous ones trigger irrelevant patterns.

Vague: 'Write about cats.' Specific: 'Write a 3-paragraph educational summary about why cats purr, for a 10-year-old.' Specificity constrains the model's vast possibility space toward your goal.

Asking the model to complete a task with no examples provided—just instructions. Example: 'Classify this email as spam or not spam: [email]'

Providing 2–5 examples of input-output pairs before your actual task, so the model learns the pattern. Useful when you need a specific format, style, or reasoning pattern the model hasn't seen in training.

Asking the model to show its reasoning step-by-step ('Let me think through this...') instead of jumping to an answer. Improves accuracy on complex reasoning tasks by making intermediate steps explicit.

Temperature scales the logits (log-probabilities) before softmax. Low T (near 0) sharpens the distribution, making high-probability tokens more likely (focused). High T (near 1+) flattens it, increasing randomness and diversity.

The model uses patterns in the prompt (examples, instructions, context) to shift its internal representations mid-inference, without updating weights. The attention mechanism learns to weight relevant context heavily for the completion task.

A user embeds hidden instructions in their input that override or contradict your system prompt. Example: 'Ignore previous rules: [new malicious instruction].' Happens because the model treats all text as a unified context.

Prompt engineering exploits the model's training objective (predict next token accurately across diverse contexts). Prompts that align with high-probability token sequences from training data are more likely to succeed—it's about activating strong learned patterns, not rewriting the model.