
Prompt engineering: writing instructions models follow well
In short: Prompt engineering is writing and testing the instructions you give a language model until it produces the output you need reliably.
Prompt engineering is the practice of designing, testing and refining the instructions given to a language model so that it produces the output you need, reliably. OpenAI's guide, read in November 2026, defines it as the process of writing effective instructions for a model. It covers stating the task clearly, giving examples, specifying format, adding context and breaking complex work into steps.
Prompt engineering in plain words
It is briefing a smart new colleague who knows a lot in general but nothing about your situation. The clearer the brief, the better the work.
How it works
Good prompts usually state:
- The role and the goal.
- The context the model cannot know.
- Constraints and format, ideally with an example.
- What to do when information is missing.
Then you test the prompt against a set of real inputs, change one thing at a time and keep what improves results. Anthropic's overview puts the testing first: before you tune a prompt, have clear success criteria and a way to test against them empirically. OpenAI's guide adds two habits for production: pin applications to a specific model snapshot, and build evals that measure the prompt when you change it or upgrade the model.
Example
Weak: "Summarise this contract."
Stronger: "You are reviewing a supplier contract for a procurement lead. List termination terms, liability caps and auto-renewal clauses as a table with clause numbers. If a term is absent, write 'not found'. Do not interpret legal risk."
Where you'll meet it
- System prompts behind AI products
- Instruction files and slash commands in coding agents
- Prompt libraries and templates in teams
- Job descriptions for AI engineering roles
What it changes for you
For everyone using AI daily, clear specification is a durable skill: it transfers across models and tools. For developers, prompts are code: version them, test them with evals, review changes. OpenAI's guide notes that reasoning models and GPT models respond to different styles of instruction, so a prompt tuned for one model needs retesting on another.
Prompt engineering vs fine-tuning
Prompt engineering changes the instructions at run time. Fine-tuning changes the model's weights with training data. Try prompts first; they are cheaper and faster to change.
Common mistake
My view: judging a prompt on three examples you picked yourself. A prompt is only as good as its results on a fixed test set that includes awkward cases. Without that, you are tuning to anecdotes.
Related terms
Context engineering · System prompt · Few-shot prompting · Chain-of-thought · Evals
From the archive: Choose prompts, RAG, tools or fine-tuning · Prompt deletion needs a regression record
A prompt is only as good as its results on a fixed test set that includes the awkward cases.


