Prompt engineering is a real skill

The same model answers the same question differently depending on how you ask. That gap is the whole job.

Part of the Getting good answers track on lAItest.

Ask a model the same question two ways and you get two different answers. Nothing about the model changed.

Only the words did.

A prompt is the entire input, not just your question.

The model receives one flat block of text: the system instructions, the conversation so far, anything you pasted, and your question at the end. It has no memory of you and no idea what you meant. It has the words. Prompt engineering is deciding which words are in that block, and in what order.

A common misconception

Commonly believed: Prompt engineering is a bag of magic phrases. Tell the model it is a world-class expert and the answer gets better.

Actually: Almost all of the reliable gain comes from boring things: stating the goal, showing an example of the output you want, saying what to do when the answer is not in the material, and being explicit about format. Magic phrases are the part that quietly stops working on the next model.

Specification, not persuasion.

Name the audience. Name the format. Give one worked example. Say what counts as a failure. Then keep the unchanging parts at the top, because prompt caching matches on a prefix and any byte you change invalidates everything after it. None of this flatters the model. All of it removes a guess it would otherwise have to make.

Your prompt works well on one model and badly on a newer one. What is the most likely cause?

Answer: The prompt quietly encoded one model’s defaults and habits. Prompts do transfer, but not cleanly. A prompt tuned hard against one model picks up small dependencies on how that model behaved: its tokenizer, its defaults, its habits. The fix is the same as the original work. Keep a handful of real test cases and re-run them whenever you switch.

In one sentence

A prompt is not a spell. It is a specification, and the parts that survive a model upgrade are the parts that state what you actually want.