Log inStart Free Trial
Meta Tech Provider

Write the prompt for your WhatsApp AI: tone, rules and what it must never promise

A good WhatsApp AI prompt is a short operating manual. It says who the business is, lists the facts the AI may use (hours, address, delivery, prices or where to find them), sets the tone and length of a chat message, and gives the exact words for handing over to a person. Anything the prompt does not cover, the AI should treat as unknown instead of guessing.

easyCoexistence runs your prompt on your own Anthropic, OpenAI or OpenRouter key in a Reply with AI step, adds the hand-off and no-invention rules under it, and pauses the AI on a chat when you answer from the WhatsApp Business app.

20,000

Characters available in the Instructions field of a Reply with AI step.

5 lines

Rules under every prompt: short plain chat text, customer's language, a person on request, no invented facts, saying it is automated when asked.

0 to 20

Earlier messages of the chat the AI reads along with the prompt; 10 by default.

600 tokens

Default Longest reply (tokens); an answer that reaches it is cut off, even mid-sentence. Adjustable from 50 to 2,000.

Escalation

Meta's messaging policy asks businesses using automation for prompt, clear and direct escalation paths to a human.

What should the Instructions field contain?

What a well-briefed new employee would need on day one, and nothing more. The Instructions of a "Reply with AI" step take up to 20,000 characters, but the whole text goes to the model with each reply, so a tight page of facts beats a pasted website. Short labeled blocks are easier to keep current than prose. Keep one source of truth: when a price, an address or an opening hour changes, change it here first, because the AI repeats whatever this field says.

Who you areBusiness name, what you sell, the city or area you serve.
Facts it may useHours, address, delivery area and fee, payment methods, lead times, policies.
Tone and lengthHow you sound, and how short a reply should be.
When it is unsureSay it will check with the team; never guess.
Hand-offWho answers, when, and the exact words to use.
Never promiseBookings, refunds, prices, stock or dates it cannot see.

What does the system add under your prompt?

Five lines, appended after your Instructions and a separator on every AI reply, tests included. The first four are listed below; the fifth makes the AI say it is the business's automated assistant when asked, and never claim to be a person, and the first AI reply to a person in a day opens by saying so. Write with them rather than against them: they already set the format, the language and a generic hand-off, so your Instructions only have to make each one specific to your business. Variables work in the Instructions too: {{first_name}}, {{name}}, {{phone}} and {{message}} are filled in before the model reads the prompt.

  • "You are replying to a customer in a phone messaging chat, on behalf of the business described above."
  • "Write only the message to send: plain text, short, no markdown headings or tables, in the customer's language."
  • "If the customer asks for a person, or you cannot help, say a person from the business will answer soon."
  • "Never invent prices, stock, dates or policies that are not in the instructions above."

How long should a WhatsApp AI reply be?

Chat length: one to three short sentences, one question back at a time, and a list only when the customer asks for options. Say so in the Instructions, because the cap is a blunt tool. "Longest reply (tokens)" stops an answer at 600 tokens by default, mid-sentence if it has to, and a WhatsApp text message holds at most 4,096 characters, so anything longer is cut. A token is a piece of a word, so leave room above the length you ask for. Put the most useful fact first, then the next step for the customer.

How do you stop the AI from inventing prices, stock or promises?

Give it the facts, tell it what it cannot see, and tell it what to say instead. The added rule already forbids inventing prices, stock, dates or policies that are missing from your Instructions; your job is to make the gaps explicit. A Reply with AI step cannot see your stock, your calendar or your orders, and it can only send text, so it cannot book, charge or refund anything; an AI agent looks things up only through the tools you give it.

  • Prices that change often: point to a live menu or price list link instead of copying numbers you will forget to update.
  • Bookings and orders: "Never confirm a booking or an order. Say that Ana confirms it in this chat."
  • Dates, discounts and refunds: "Never promise a delivery date, a discount or a refund."
  • Health, legal or money questions: "Do not give medical advice. Offer to book a consultation."
  • Live data such as order status belongs in your own system, reached through a Call my webhook or Call an API step that returns the real value, or through a custom tool on an AI agent.

How should the hand-off to a person work?

Meta's messaging policy allows automated replies inside the 24-hour window when people also get a prompt, clear and direct path to a human, such as an in-chat transfer, a phone number or an email. On a Coexistence number the in-chat transfer is natural: the person answers in the same chat from the WhatsApp Business app, and the automation pauses on that chat for 4 hours by default. A Hand to a person step does the same from the flow: it pauses the bot on that chat, 24 hours by default, and emails you. Make the hand-off concrete in the Instructions: who answers, when, and the exact sentence to send, such as "I'll pass this to Ana now. She answers here between 7:00 and 19:00." List the triggers too: a request for a person, a complaint, a change to an order, anything involving money.

What does a complete WhatsApp AI prompt look like?

Like the one below, written for a fictional bakery in the blocks described above. It works as is in a Reply with AI step: swap in your own facts and menu link, keep the structure, and keep the never-rules even when they feel obvious, because they are what stops a confident wrong answer. Test it with hard questions before switching the automation on.

Instructions for a Reply with AI step (fictional example)
You are the WhatsApp assistant of Ana's Bakery in Campinas, Brazil.
You answer customers' questions about our products, hours, orders and delivery.

FACTS
- Open Tuesday to Sunday, 7:00 to 19:00. Closed on Mondays.
- Address: Rua das Flores 120, Cambuí. Pickup at the counter.
- Delivery within 5 km, fee R$ 8, from 9:00 to 18:00.
- Cakes need 48 hours' notice. Gluten-free bread on Tuesdays and Fridays.
- Product prices change often: send the menu link https://example.com/menu instead of quoting them.
- Payment: Pix, or card on delivery.

STYLE
- Warm and brief: one to three short sentences, one question at a time.
- Use {{first_name}} only in your first reply.
- No lists unless the customer asks for options.

WHEN UNSURE
- If the answer is not in FACTS, say you will check with the team.
- If the message is only [voice message] or [image] with no words after it, ask the customer to type the question.

NEVER
- Never confirm an order, a delivery time, a discount or a refund.
- Never say you are a person. If asked, say you are the bakery's virtual assistant.

HAND-OFF
- If the customer asks for a person, complains or wants to change an order, reply:
  "I'll pass this to Ana now. She answers here between 7:00 and 19:00."

What should you test before the AI goes live?

The questions that break weak prompts, not the friendly ones. A test_automation call reads only the sample you type, without earlier messages, so put any context into the sample itself; the editor's Try it as a customer keeps the turns of its own chat. Read each answer as the customer would: short enough, right facts, right tone, and the hand-off sentence when it should appear. The token counts on each test show what a reply costs. Then tune "Earlier messages it reads": 0 gives no memory, 10 is the default, and more helps long conversations at a higher cost per reply.

  • A price or product that is not in the prompt: it should send the menu link or say it will check.
  • An angry message about a late order: it should apologize briefly and hand over to a person.
  • "Can I talk to someone?": it should send your exact hand-off sentence.
  • A question in another language: it should answer in that language.
  • Something off-topic, such as a request to write an essay: it should steer back to the business.

How to set it up

  1. Open the Reply with AI stepOn the Automations tab, open your automation and click the "Reply with AI" block in the diagram.
  2. Write the InstructionsFill in the blocks: who you are, facts, style, when unsure, never, hand-off. Keep to what customers actually ask about.
  3. Set the two numbers"Earlier messages it reads" starts at 10, and "Longest reply (tokens)" at 600, which leaves room for a short answer.
  4. Save with the switch Off and run testsPress Save, open Test, and use "Run test" with five hard questions; nothing is sent while you test.
  5. Switch it On and read the first runsTurn it On, press Save, and read the first real replies in the Runs tab, then adjust the Instructions where an answer read wrong.

Common mistakes

  • Pasting a whole website or catalog into the Instructions. The full prompt goes to the model with each reply, so a long one costs tokens on each AI reply; keep what customers ask about and link the rest.
  • Copying prices that change weekly into the prompt and forgetting them. The AI quotes what the Instructions say; link a live price list, or answer price questions with a keyword reply you update in one place.
  • Asking for long, formatted answers. The added rules discourage headings and tables, and an answer that reaches "Longest reply (tokens)" stops mid-sentence.
  • Testing only friendly questions. A prompt that handles "what time do you open?" can still invent a price or argue with an angry customer, so test the hard cases before switching it On.
Doing this with EasyCoexistence

Try it runs the prompt on a sample without sending anything to anyone. US$ 9 down to US$ 0.75 per number per month, falling as you add numbers, with the first 3 days free.

easyCoexistence is a multichannel API and MCP server that connects AI agents such as Claude and ChatGPT to WhatsApp, Instagram, TikTok, Facebook and more. On WhatsApp it uses Coexistence: the business keeps using its number in the WhatsApp Business app while the AI works through Meta's official API.

Frequently asked questions

Should the prompt be in English or in my customers' language?

Either works. The added rules make the model reply in the customer's language, so write the Instructions in the language you write most precisely, and include the local words customers use for payment, delivery or products.

Can the AI read my website or catalog by itself?

Not in a Reply with AI step, which reads only its Instructions and the chat, with no browsing. An AI agent can take uploaded files, URLs and text as knowledge. For a Reply with AI step, paste the facts customers ask about, link the rest, and hand questions that need live data to your own system with a Call my webhook or Call an API step.

Can I use different prompts for different questions?

Yes. Create separate automations with keyword triggers, each with its own Reply with AI step and Instructions, plus a catch-all for everything else. Keyword automations answer first, and a message gets at most one automation.

Does the AI remember a customer from last week?

Only through the chat history it reads: up to 20 earlier messages for a Reply with AI step (10 by default) and up to 40 for an agent (20 by default), kept for the days you choose, 90 by default. There is no long-term memory in the step, but contacts keep fields and tags, and an agent keeps notes on a contact. Facts that must persist, such as a customer's usual order, belong in a contact field or your own system.

Keep reading

Ready to get started?

Set up WhatsApp Coexistence in minutes, not months. The app keeps working on the phone.

Start Free Trial

Verified on