Prompt engineering
Prompt engineering is the practice of writing instructions that get reliable output from a language model. It covers what you ask for, how the request is structured, what context travels with it, and which examples are included to demonstrate the behaviour you want.
Prompt engineering was treated as a discipline in its own right when models needed careful handling to perform well. Newer models follow plainer instructions, so the craft has shifted from clever phrasing towards supplying the right context and constraints. It is not a safety mechanism, and that is the misunderstanding which matters most in regulated settings.
How does prompt engineering work?
A prompt usually carries four things: the role the model should take, the task, the context it needs, and the constraints on its output. Most quality problems trace to the third and fourth being thin rather than the first two being badly worded. Techniques that still earn their place:
Worked examples: showing two or three cases handled correctly, rather than describing the rule
Explicit constraints: naming what must not appear, not only what should
Structured output: requesting a defined format when something downstream will parse it
Room to reason: letting the model work through a multi-step problem before answering
What has faded is phrasing trickery. Contemporary models respond to clear instructions, so effort spent on elaborate wording pays back less than effort spent on giving the model what it needs to be right.
Why is a prompt not a control in a regulated business?
A prompt is a request, and the model usually complies. Usually is the problem. A prompt telling a model never to promise a refund will hold most of the time, will not hold every time, and nothing in the architecture makes it hold.
A control has to inspect what actually happened. The difference in practice:
A prompt asks the model to behave a certain way, before it responds
A guardrail examines what the model produced, and intervenes before the customer sees it
That gap is why an evidence trail cannot be built on prompts alone. Prompt engineering makes good behaviour more likely, and a regulator asking how you prevent something wants a check rather than an instruction. Our guide on AI copilots versus autonomous agents covers where those checks sit.
Frequently asked questions
Is prompt engineering still a useful skill?
Yes, though it looks different from two years ago. The value has moved from finding phrasings that unlock a model towards structuring context, constraints, and examples well. The underlying skill is specifying a task precisely, which was always the valuable part.
Can prompt engineering replace guardrails?
No. A prompt shapes what the model attempts, and a guardrail checks what it produced. Prompts fail silently and occasionally, which is exactly the failure mode a control exists to catch. Regulated deployments need both, doing different jobs.
Who should own prompts in a financial services deployment?
Whoever owns the policy they encode, which is usually operations or compliance rather than engineering. Where changing a prompt requires an engineering release, every policy change joins a queue, and that delay is the practical problem most teams hit after go-live.