How to write instructions (a system prompt) for your AI assistant

Imagine you hire someone to handle your business and you tell them only 'answer the messages'. They will most likely reply however they feel like: sometimes well, sometimes with the wrong price, sometimes in a way that does not match your brand. Now imagine that instead you hand them a sheet with clear instructions on who they are, what your business offers, how to speak, and what they must never promise. That, applied to an AI, is a system prompt.

What a system prompt is

A system prompt is the base set of instructions an AI assistant receives before talking to anyone. It defines who it is, who it serves, what information it handles, how it should reply, and what limits it has. While each customer writes different messages, the system prompt is fixed: it is the manual the assistant always keeps in mind, conversation after conversation.

It is, without exaggeration, what most influences the quality of the replies. Writing good instructions can noticeably improve what the assistant produces and reduce invented or off-topic answers. An assistant without a good system prompt is like an employee without training.

Instructions should use simple, direct language and contain the minimal set of information that fully outlines the behavior you expect.

The pieces of a good system prompt

The guides from the companies that build these models agree on a few essential blocks. You do not need all of them to start, but keeping them in mind helps:

• Identity: who the assistant is and who it serves. For example, 'You are the assistant of a barbershop and you serve customers who want to book a haircut'.
• Business information: services, prices, hours, location, and policies. What the assistant needs to know to reply well.
• Tone and style: how it should sound, formal or friendly, and in which language.
• Rules and limits: what it must never do, for example promise discounts that do not exist or give medical advice.
• What to do when it does not know: who to hand off to, or how to ask for help instead of making things up.

Write clearly and at the right altitude

The most common mistake is going to an extreme. On one side are instructions that are too vague, like 'be friendly and help', which give the assistant no concrete signals about what to do. On the other are instructions so rigid and full of rules that they become brittle and hard to maintain. The sweet spot is in between: specific enough to guide behavior, yet flexible enough to give the model strong judgment.

Use simple, direct language, as if explaining to a new person. Minimal does not mean short: it means including all the necessary information, without filler. And forget phrases like 'please follow these instructions very carefully'; they do not help, they just take up space.

Examples are worth more than a thousand rules

One of the most useful techniques is showing examples. Instead of describing in the abstract how you want it to reply, write out a couple of real cases with the ideal answer. For a model, examples are like pictures worth a thousand words. If you want it to share prices a certain way, show it exactly how.

• Pick typical situations: someone asks the price, someone wants to book, someone arrives upset.
• Write the ideal reply for each, with the right tone and the right information.
• Include a tricky case too, so it knows how to react.

Improve it through practice

Do not expect to write the perfect system prompt on the first try. The experts' recommendation is to start with a minimal, clear version, test it with real cases, and fix it based on the mistakes that show up. Every time the assistant replies with something you do not like, adjust an instruction. After a few rounds, it gets well tuned.

When you use an assistant like Lidia, much of this work is done by us and adapted to your niche, but understanding what a system prompt is helps you give us the right information about your business, which is exactly what feeds those instructions.

The takeaway

The system prompt is your AI assistant's manual: it tells it who it is, what it knows, how to speak, and what it must not do. Write it in simple language, give it your business information, set clear limits, and teach it with examples. Start simple and improve it through practice. The quality of your assistant will never be better than the quality of the instructions you give it.

Sources

• Anthropic — https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents
• OpenAI — https://help.openai.com/en/articles/6654000-best-practices-for-prompt-engineering-with-the-openai-api
• Agent Wiki — https://agentwiki.org/how_to_structure_system_prompts