The technique ladder: simple first
Prompt engineering is a ladder; each rung is only worth it when the one below failed:
- Zero-shot: just the instruction. For everyday tasks on strong models, this works immediately.
- Few-shot: two or three input/output examples. For subtle formats and specific styles.
- Chain-of-Thought: asking for step-by-step reasoning before the answer. For multi-step math and logic.
- System prompts: keep fixed rules (role, constraints, format) separate from user input at the API level — more stable and reusable.
- RAG: when the answer depends on knowledge the model does not have.
- ReAct: combining reasoning and action — the model thinks, calls a tool, reads the result and continues.
The practical selection rule
Start at the bottom of the ladder and move up only with evidence. If zero-shot is right 90% of the time, do not add few-shot; unjustified complexity hurts maintainability and cost.
What no longer works
- "This is very important for my career" and similar emotional appeals have no meaningful effect on current models — drop them.
- Repeating the same instruction ten times; one clear sentence beats ten redundant ones.
- Copied "magic prompts"; those were tuned for one model or version and do not generalise to your task without testing.
Treat prompts like code
Keep important prompts in git, record every change, and run a small evaluation set on each change. A prompt without tests breaks the moment you switch models or versions.
