How to write the instructions for your AI agent so it answers well

You hire a new sales rep and on day one you only tell them: “Handle the WhatsApp messages”. Without explaining the products, the prices, how to greet people, what to do if someone complains or when to pass the case to a supervisor. How well do you think they will do?

Exactly the same happens with an AI agent. The model can be brilliant, but without good instructions it improvises. The good news: writing effective instructions has a technique, and Anthropic, the company behind Claude, documents it in detail.

Why do instructions matter so much?

Because they are your agent's training manual. According to Anthropic's prompt engineering overview, adjusting the prompt is much faster than other methods of controlling model behavior and often delivers big improvements in little time.

Before writing, be clear about what your agent does; if you do not have one yet, read what an AI agent for WhatsApp is.

Which techniques does Anthropic recommend?

Five, ordered from the most broadly effective to the most specific:

Technique	What it means for your agent
Be clear and direct	Say exactly what to do, without assuming it will guess
Give examples	Show two or three model conversations
Let it reason	Ask it to think before answering in complex cases
Use XML tags	Separate instructions, data and examples with clear markers
Assign a role	Define who it is: “sales advisor for a clinic in Bogotá”

According to Anthropic's best practices guide, XML tags help the model interpret prompts that mix instructions, context, examples and variable data without ambiguity.

What does a good set of instructions look like?

A practical structure for a WhatsApp agent:

1. Role: who it is and who it represents.
2. Goal: what it must achieve (answer questions, qualify, book).
3. Business information: products, current prices, hours, coverage areas.
4. Rules with their reason: “do not promise delivery dates, because they depend on the day's inventory”.
5. When to hand over to a person: complaints, payments, cases outside its information.
6. Examples: two or three well-handled conversations.
7. Format: short messages, a single message per reply, informal tone.

Common mistake: writing only prohibitions (“do not do this, do not say that”). It works better to say what to do instead and explain why: that way the agent applies judgment in cases you did not foresee.

Why ask for answers in a single message?

For customer experience and for cost. An agent that splits each idea into five bubbles tires the customer and, since October 2026, every reply through the WhatsApp API counts toward charges; we explain it in WhatsApp's new charges.

How do I know if my instructions work?

By testing them with real conversations and fixing what fails.

• Review a sample of conversations every week.
• Note where it answered badly, made things up or did not hand over the case in time.
• Adjust the instruction or add an example that covers that case.
• Keep the business information up to date: an old price in the instructions is an old price in every answer.

Watch out: instructions do not replace data. If the agent needs to check inventory or availability in real time, it must connect to those systems, for example with MCP or with n8n workflows.

In short

An AI agent is only as good as its instructions: a clear role, a goal, business information, explained rules, examples and a defined format. It is the same onboarding you would give a new sales rep, written so the AI follows it every time. We design, test and fine-tune your agent's instructions in our AI agents service.