What is prompt engineering?
A demo works on the wording one person found by accident, and nobody can say why the same request fails for a colleague.
Prompt engineering
prompting
Prompt engineering is the practice of writing and refining the instructions sent to a language model so it produces the output you want consistently. The work is choosing what to say, what to show as an example, and what context to include.
The techniques are few and boring, which is a good sign. Give the model a role and clear instructions. Show a handful of worked examples. OpenAI calls that few-shot learning, and describes it as steering a model toward a new task with a few input and output pairs. Mark the boundaries of your content too, which OpenAI suggests doing with a mix of Markdown formatting and XML tags.
None of that is a secret phrase. Most of the gain comes from saying plainly what you want and supplying the facts the model needs to do it.
Newer models want less instruction, not more
The advice has reversed for reasoning models. OpenAI's own guidance states that reasoning models "will provide better results on tasks with only high-level guidance", and that this differs from GPT models, which "benefit from very precise instructions". Its analogy is a senior co-worker against a junior one.
Read what that does to the job. Elaborate step-by-step scaffolding, written for an older model, can now make results worse rather than better. The craft has not vanished. It has moved to deciding which documents reach the model, which tools it may call, and how you know the answer was right.
Treat prompts as code either way. Keep them in the repository, change one thing at a time, and score the change against saved examples.
- InstructionShorter than it used to be.
- ContextThe right passages, not all of them.
- ToolsWhat the model may call.
- Output shapeA schema, so code can read it.
- ScoreSaved cases, run on every change.
Four of these five survive a model upgrade. The wording is the part that does not, which is why it is the weakest place to invest.
Related questions
01Is prompt engineering a dying skill?
The narrow version is fading and the broader one is not. Hunting for magic wording is worth less every release, since models increasingly do well on plain instructions. Deciding what context, tools and checks a system needs is the same skill grown up, and that work is increasing.
02What is the difference between prompt engineering and context engineering?
Prompt engineering is about the wording of the instruction; context engineering is about what information accompanies it. The second matters more in production, because a well-worded request over the wrong documents still returns a wrong answer.
03Do we still need examples in the prompt?
Yes, when the output has a shape that is easier to show than to describe. Two or three examples fix formatting, tone and edge-case handling faster than a paragraph of rules. Drop them once a schema and a scored test set are doing that job.
