Prompt Engineering Primer

# Prompt engineering primer

> Six patterns, one example each, all from job-radar. Read top-to-bottom once. After that, use it as a reference card while writing your first prompts.

---

## The two-minute mental model

LLMs are next-token predictors. Everything in the prompt — instructions, examples, format hints, role framing — biases that next-token prediction. The art is making the bias point at the answer you want **for inputs you haven't yet seen.**

The single biggest beginner mistake: writing a prompt that works on the input you tested, then assuming it generalizes. It usually doesn't. Generalization comes from:

1. **Structure** (system role, clear sections, structured outputs)
2. **Examples** (few-shot — the model copies patterns)
3. **Constraints** (enums, schemas, "if X then Y" rules)
4. **Iteration on diverse data** (this is what evals are for)

---

## Pattern 1 — System prompt vs user prompt

**Anthropic + OpenAI both split prompts into system (sets behavior) and user (carries the task).**

System content sticks across turns. User content is the per-call payload. Putting your role/persona/style guide in the system prompt and the task-specific text in user means your prompt is reusable.