The core five, and what each actually buys
System prompts set durable instructions that outrank the conversational turn — the place for role, rules, and refusals. Few-shot examples show rather than tell; for output shape and edge-case handling they remain the highest-yield lines in any prompt. Chain-of-thought — asking for reasoning before the conclusion — trades tokens for accuracy on problems with intermediate steps. Structured output constrains responses to a schema software can parse, turning prose into data. And placement matters more than prose style: models weight the beginning and end of context most heavily, so burying the critical instruction in the middle of a long prompt is a self-inflicted wound.
Advanced means composed, not exotic
Most techniques sold as advanced are the core five arranged with intent. Decomposition splits a hard task into chained simple ones, each with its own focused prompt. Self-critique runs a second pass that reviews the first against criteria you state. Role and audience framing sharpens register more reliably than adjectives. Template discipline — versioned prompts with variables, tested like code — outperforms heroic one-off prompts, because the prompt that cannot be reproduced cannot be improved. The pattern across all of it: clarity, examples, and structure beat clever wording every time it matters.
Testing: the technique nobody lists
A prompt change is a behaviour change shipped to production, and the teams that treat it that way pull ahead. Keep a small evaluation set of real inputs with known-good outputs; run it on every prompt edit and every model-version change, because providers update models under stable names and yesterday's prompt quietly degrades. This is the same evaluation discipline [agents require](/guides/build-an-agentic-ai-system), applied one level down — and it is what separates prompt engineering from prompt guessing.
The boundary that no technique crosses
Every technique here shapes likely behaviour; none guarantees it. A sufficiently adversarial input can argue a model out of its instructions — which is why prompts are advice and [the tool surface, permissions, and runtime gates](/guides/secure-agentic-ai) are the boundary. In the agent era the craft also widens into context engineering: tool descriptions, retrieved documents, and accumulated history are all prompt surface, much of it written by other people. Treat the prompt as the steering wheel, never the brakes.