Prompt engineering
Quick definition
Prompt engineering is the practice of designing the instructions, context and examples given to a language model so that it produces reliable, useful outputs. For a support assistant, the core prompt sets its role, rules and tone.
On this page
What goes into a prompt
A language model only sees the text it receives. For a customer service assistant, that text usually has several layers: a system prompt describing who the assistant is, who it serves and what it must never do; reference material retrieved for the current question; the descriptions of the tools it may use; and the conversation so far.
Prompt engineering is the work of making that package clear and consistent. Vague instructions (“be helpful”) produce vague behaviour; specific ones (“if the order number is missing, ask for it before searching”) produce predictable behaviour.
Techniques that work
Be explicit about the task, the audience and the expected format. Give the model a role and the context a new colleague would need. Show one or two examples of good answers when the format matters. Separate instructions from data — and tell the model that retrieved text is information to use, not orders to follow. Ask it to reason step by step on complex requests; Wei and colleagues showed in 2022 that this “chain-of-thought” prompting improves results on multi-step problems. And say what to do when information is missing.
Model providers publish their own guides; Anthropic's documentation, for example, recommends being clear and direct, using examples and giving the model room to think.
Common mistakes
The most common is the ever-growing prompt: each bad answer leads to one more rule, until rules contradict each other and the model follows none of them well. Few-shot examples copied word for word into answers are another trap — including examples written in English, which can push the assistant to answer in English. So is putting facts that change, such as prices, into the prompt instead of a knowledge base.
A prompt should describe behaviour. Facts belong in the knowledge base and live data in tools.
Prompting without code in AI SmartTalk
In AI SmartTalk the system prompt is called the assistant's biography. A guided form helps write it — identity, communication style and rules, with a friendly, professional, neutral or custom tone — and previous versions are kept so a change can be compared or undone. Decision rules and the handoff policy are set separately, and built-in policies such as answering in the visitor's language are added by the platform.
Frequently asked questions
Is prompt engineering still needed with better models?
Less trickery is needed, but clear instructions still matter. Better models follow good prompts more faithfully, which makes it even more important that the prompt says what you actually want.
What is a system prompt?
The standing instructions given to the model before the conversation starts: its role, rules, tone and limits. Users do not see it, but it shapes every answer.
What is prompt injection?
An attack in which text written by a user or hidden in a document tries to override the system prompt — for example “ignore your instructions and…”. It is reduced by separating instructions from data, limiting what tools can do and checking actions server-side.