WACrew

Blog

How to make AI write like you on WhatsApp

A method to make AI write messages in your own style, measured on your own conversations, and to catch the habits that give AI text away before it is sent.

Updated 5 min read
On this page
  1. Why "be more natural" doesn't work
  2. Step 1: collect your messages
  3. Step 2: measure what you do
  4. Step 3: write a style guide with examples
  5. Step 4: catch the tells before sending
  6. Step 5: deliver at a human pace
  7. How WACrew does it
  8. FAQ

Asking an AI to "sound more human" almost never works. It swaps "Dear customer" for "Hi there! 😊", adds an "I completely understand" and still sounds like an automated help desk. The reason is simple: "human" is not a style. You write in a specific way, and that way is what your clients recognize.

This guide shows how to make AI write like you for real. You measure your style on your own messages, turn the measurements into rules a model can follow, and check the text before it goes out.

Why "be more natural" doesn't work

A language model learns from a huge average of text. When you ask for natural, it gives you the average of what usually gets called natural in customer service: an upbeat greeting, emoji, textbook empathy and a closing offer to help with anything else. That is exactly what clients have already seen in every chatbot.

The fix is the opposite of a generic instruction. Instead of telling the model how not to write, you show it how you write, with concrete numbers.

Step 1: collect your messages

Gather a few hundred messages you wrote to clients, ideally from before you used any automation. Export chats from WhatsApp (the app has an Export chat option) and keep only the lines that are yours.

Two precautions:

  • Separate registers. Nobody writes the same way to clients and to friends. Mix them and the AI will bring your group-chat slang into a conversation with a client's director.
  • Strip personal data before pasting anything into a tool: phone numbers, IDs, addresses, contract amounts.

Step 2: measure what you do

Read your messages looking for patterns you can count. These carry the most weight:

TraitWhat to measure
LengthHow many characters a typical message of yours has. Do you send short bubbles in a row, or one paragraph?
Final punctuationWhat share of your messages ends with a period. Many people almost never use one in chat.
OpeningsHow you start. "Hey", "Morning", the person's name, or straight to the point?
Abbreviations"thx" or "thanks", "u" or "you", "pls" or "please"
EmojiHow often, and which ones
Form of addressFirst name, nickname, no name at all?

The result is ranges, not fixed rules. If you end one message in ten with a period, the AI should do it rarely too, not never. Text where every bubble looks the same also gives a machine away.

Step 3: write a style guide with examples

Turn the measurements into a short guide the model reads before writing. The format that works has three parts.

Measured ranges. "Short messages, usually around two lines. Periods are rare. Opens with 'hey' or goes straight to the point."

Real examples. Five to ten real messages of yours in different situations: confirming something, giving bad news, chasing an answer, turning down an out-of-scope request. One example is worth more than any adjective.

Positive instructions. Say what to do, not what to avoid. "Start with the answer" works better than "don't open by praising the question". A list of forbidden phrases tends to plant the very phrase you banned.

Step 4: catch the tells before sending

Even with a good guide, the model slips. That is why an automated review that runs after writing and before sending is worth it. It looks for the classic signs:

  1. an opening that praises the question ("Great question!")
  2. a reflex closing offering help with anything else
  3. em dashes linking clauses
  4. adjective triplets and punchlines
  5. repeated apologies and over-the-top self-blame
  6. the same opening as your last few messages

This filter can be plain code with fixed rules. It is fast, free and catches most cases. The open-source library botsmell does exactly that, and for each hit it returns a rewrite suggestion you can feed back to the model. The full list of signs is in the tells that make AI text read like a bot.

Step 5: deliver at a human pace

A perfect message that arrives two seconds after the client's question still feels like a bot. People read, think and type. If delivery respects a reading delay, shows "typing" for the right amount of time and splits the text into bubbles the way you would, the client never stops to wonder who wrote it.

How WACrew does it

WACrew applies these five steps by default. The AI writes from a voice skill measured on your own messages, with ranges and examples, and reads the conversation history before replying. A code-based review blocks the tells listed above, compares the draft with the account's recent messages and rejects it if it stands out too much; the AI then rewrites before anything leaves. Delivery has read receipts, typing and pauses, and it stops on its own if the client writes again.

Nothing is sent without you seeing it, unless you switch that specific chat to autonomous mode.

FAQ

How many messages do I need to measure my style? A few hundred already show a pattern. Going beyond a few thousand rarely changes the ranges.

Can my whole team share one style guide? It is not ideal. Each person writes differently, and a client who has talked to the same person for months notices the switch. If it is a company account, define a house voice and measure it on the messages of whoever answers the most.

Will the AI pretend to be me? It writes in your style, from your account. If a client asks whether they are talking to an AI, the right answer is the truth, and ideally that question goes to you.