Glossary
Prompt
A prompt is the instruction in natural language that tells a language model how to behave and what to do.
On this page
For a phone assistant, the system prompt defines the role: how the assistant introduces itself, which tone it strikes, which questions it asks, what it must not promise and when it hands over to people. It is written once and applies to every call. What the caller says is something different; that is the input during the ongoing conversation.
Good prompts are specific and short. They describe procedures step by step, give examples for borderline cases and contain no contradictory rules. For the phone there is more: short sentences, one question per turn, numbers and times worded so they can be spoken. Changes to the prompt take effect on all calls immediately and should therefore be checked with test calls before going live.
In practice
A usable prompt for the phone follows the course of a call: greet, clarify the request, give information or take details, summarise, say goodbye. For each step, write what the assistant should do and what it needs for it. Prohibitions work best together with the alternative: not just “do not quote prices” but “for price questions, offer a callback with a quote”. After every change, make the same five to ten test calls again. That is how you notice whether a new rule has shifted something elsewhere.
An example
A restaurant writes in its prompt: “Take reservations in a friendly manner.” In testing, the assistant books a table for twelve on a Saturday evening, although groups of eight or more are only possible by arrangement. The revised instruction reads: “Ask for date, time and number of guests. Up to seven guests, reserve directly. From eight guests, take name and callback number and say that the team will be in touch.” After that the assistant behaves as intended in both cases.
A common misunderstanding
Prompts grow longer over time, because another rule is added after every mistake. At some point two of them contradict each other, and the assistant follows now one, now the other. Cutting is therefore just as important as adding. A second misconception is that facts belong in the prompt. Opening hours, prices and contacts change and are better kept in the knowledge base. The prompt describes behaviour; the knowledge supplies the content.
On this topic:
As of: 12 October 2026