How to Write Great AI Agent Instructions for WhatsApp

The instructions you give your agent decide everything: its tone, its limits, when it answers, and when it steps aside. Good instructions feel like a great employee onboarding. Vague ones lead to wandering, off-brand replies. Here is how to write instructions that work.

Start with a clear persona

Tell the agent exactly who it is and how it speaks. Be specific. “Be helpful” means nothing. “You are a calm, warm support assistant who keeps replies short and never uses slang” means a lot.

Cover three things:

Identity — the business name and the agent’s role.
Tone — warm, formal, playful. Match your brand voice.
Goal — what a successful conversation looks like.

Set firm rules and boundaries

This is where most of the value lives. Tell the agent what it must never do. Customers will push, and clear boundaries keep the agent honest.

Never invent prices, dates, or facts you have not given it.
Never promise refunds, discounts, or outcomes you have not approved.
If it does not know, say so and offer to connect a person.

An agent that admits “I am not sure, let me get a teammate” builds far more trust than one that guesses confidently and gets it wrong.

Tell it when to hand off

The handoff to a human is one of your strongest tools. Spell out exactly when to use it:

The customer asks for a human.
The customer is upset or the situation is sensitive.
The request is outside what the agent can confidently handle.

When you name these triggers clearly, the agent stops guessing and pulls in a teammate at the right moment. The teammate gets notified and takes over from the shared inbox.

Lean on Skills for the details

Resist the urge to cram every policy into the instructions. A bloated prompt makes the agent slower to follow and easier to confuse.

Instead, keep the instructions focused on persona and behavior, and move detailed playbooks into Skills. Your returns process, your booking steps, your shipping rules each become a Skill the agent loads only when it is relevant. Your core instructions stay clean, and your answers stay accurate.

Give examples

Models learn from examples fast. If you want a certain style, show it.

A line like this does a lot of work: “If a customer asks about hours, reply like this: ‘We are open 9am to 8pm, Saturday through Thursday. See you soon!’” One good example shapes dozens of future replies.

A short example persona

Here is a complete, compact persona you can adapt:

You are Nour, the friendly support assistant for Layla Bakery. You speak warmly and keep replies short and clear. You help customers with opening hours, location, and custom cake inquiries. You only state facts you have been given. You never quote prices or promise delivery times you are unsure of. If a customer is upset, asks for a person, or needs something you cannot confirm, hand off to a human teammate right away. Always reply in the customer’s language.

That is enough to produce sharp, on-brand replies, with the heavy details living in Skills.

Common mistakes to avoid

Too vague. “Be nice and helpful” gives the agent nothing to stand on.
Too bloated. A giant prompt stuffed with every policy. Use Skills instead.
No boundaries. Without firm limits, the agent will improvise, and improvising means guessing.
No handoff rules. The agent tries to solve everything alone and frustrates customers it should have passed on.
Never testing. Write, test in the Playground, refine. Repeat until it feels right.

Great instructions are short, specific, and honest about limits. Get the persona and boundaries right, let Skills carry the details, and your agent will sound like your best teammate, every single time.