Prompt engineering

What actually moves the result

Most of the gain comes from a small number of unglamorous moves:

• Say what the output must look like. "Return a JSON object with the keys title and steps" is worth more than any amount of politeness. Most APIs also support a schema-constrained mode, which is stronger than asking.
• Give context instead of relying on recall. Paste the actual document, the actual error, the actual code. This is the single largest source of improvement and it also cuts hallucination.
• Show two or three examples of input and desired output. Few-shot examples communicate tone and edge cases faster than a paragraph of description.
• State the constraints and the failure case. What to do when the answer is not in the material, what to leave out, what not to invent.
• Split the work. One model call per decision, chained, beats one call asked to do five things — and it gives you somewhere to inspect when it goes wrong.
• Put the critical instruction near the end of a long prompt, where attention is strongest.

For anything you will run more than a handful of times, keep a small set of test inputs and compare versions of the prompt against them. Without that you are guessing, because a change that fixes one example routinely breaks two others.

What to watch out for

Prompts are not portable, and they are not a substitute for design. A prompt tuned on one model often performs worse on another; changing model version is a reason to re-test, not a drop-in swap. Newer reasoning-oriented models have also made some old tricks obsolete or harmful — telling a model to "think step by step" when it already reasons internally can be noise, and elaborate role-play framing usually adds nothing over a plain, precise instruction.

The wider point: prompting cannot fix a problem the model genuinely cannot do. If the task needs current data, retrieve it. If it needs a private codebase, supply it. If it needs a fixed behaviour across thousands of calls, fine-tuning or program structure is the answer. Rewriting the prompt a twelfth time is how people avoid noticing that.