The Generic AI Problem

You've set up AI auto-replies. They're fast. They're helpful. But they sound like... well, like AI.

"Thank you for reaching out! I'd be happy to assist you with that inquiry. Please allow me to provide you with the information you requested."

Nobody talks like that. It's technically correct but soulless. Customers can tell immediately that they're talking to a bot, and trust plummets. The words are right but the voice is wrong.

What if your AI could write like you? Not just follow your prompts, but actually mimic how you structure sentences, the words you choose, the way you use punctuation and emojis. That's what Learn Mode does—it trains on your actual message history and generates responses that sound like they came from you.

This guide explains how style learning works, how to train a style profile, and when to use it versus custom prompts.

What Learn Mode Analyzes

Learn Mode isn't magic—it's pattern recognition applied to your writing. Here's what the AI extracts from your message history:

Sentence length patterns:

Do you write short, punchy sentences? Or longer, flowing ones?
Average word count per message
Variation in message length (some people are consistent, others swing between one-word replies and paragraphs)

Vocabulary fingerprint:

Common words you use that others don't (industry jargon, regional slang)
Words you avoid (formal vs casual)
Technical terms vs plain language

Emoji and punctuation habits:

Do you use emojis liberally or sparingly?
Which emojis do you prefer? (🔥 vs 😊 vs 👍)
Punctuation style: periods, exclamation points, ellipses, or none?

Tone markers:

Formality level (Hey vs Hello vs Hi)
Question style (You good? vs How are you? vs I hope you're doing well)
Urgency indicators (ASAP, urgent, when you get a chance)

Example comparison:

Without Learn Mode: "Hello! Thank you for your interest in our services. I would be pleased to schedule a consultation call with you at your earliest convenience. Please let me know what times work best for your schedule."

With Learn Mode (trained on casual user): "hey! yeah let's do a call. i'm free tues/wed afternoons, lmk what works 👍"

Same information. Completely different voice. One sounds like a corporate email from 2005. The other sounds like a real person texting a friend.

Step 1: Prepare Training Data

Quality in, quality out. Learn Mode needs at least 50 messages to build a reliable style profile, but 200+ is ideal.

Start a training job:

bash

curl -X POST https://apiv2.waiflow.app/api/v2/ai/style/train \
  -H "Authorization: Bearer YOUR_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "contact_id": null
  }'

Parameters:

contact_id: Optional. Set to null to train on all your outbound messages, or specify a contact JID to train on how you write to a specific person/context.

When to use contact-specific training:

You write differently to customers vs colleagues
You have distinct personas (sales vs support)
Different products need different tones

Most users start with a global style profile (contact_id: null) and add contact-specific ones later if needed.

Data quality matters:

Good training data:

Actual business conversations (not personal chit-chat)
Messages you wrote, not copy-pasted templates
Variety of topics and contexts
Mix of questions, answers, and proactive messages

Bad training data:

Generic auto-responses ("Thanks for your order!")
One-word replies ("ok", "yes", "thanks")
Messages from other people in a shared account
Error messages or system notifications

If your message history is polluted with low-quality data, clean it first:

Export messages via API
Filter out automated/template messages
Keep only human-written content
Re-import as training corpus

Step 2: Train the Model

Training takes 2-5 minutes depending on message volume. MoltFlow uses a Celery background worker to process your message history and extract style features.

Check training status:

bash

curl -X GET https://apiv2.waiflow.app/api/v2/ai/style/status/{task_id} \
  -H "Authorization: Bearer YOUR_API_TOKEN"

Response during training:

json

{
  "task_id": "abc123",
  "status": "processing",
  "progress": 67,
  "estimated_completion": "2026-02-14T15:35:00Z"
}

Response when complete:

json

{
  "task_id": "abc123",
  "status": "completed",
  "profile_id": "profile_xyz789",
  "sample_count": 487,
  "features_extracted": 23
}

What's happening under the hood:

System loads your message history from the database
Text preprocessing: tokenization, normalization, emoji extraction
Feature extraction: sentence stats, vocabulary frequency, n-gram patterns
Style embedding creation: compressed vector representation of your writing style
Profile saved to database for runtime use

The style embedding is approximately 1KB—compact enough to load on every AI generation request with zero latency impact.

Step 3: Review Style Profile

Once training completes, inspect your style profile to verify it captured your voice:

bash

curl -X GET https://apiv2.waiflow.app/api/v2/ai/style/profile \
  -H "Authorization: Bearer YOUR_API_TOKEN"

Example response:

json

{
  "id": "profile_xyz789",
  "tenant_id": "tenant_abc",
  "contact_id": null,
  "features": {
    "avg_sentence_length": 8.3,
    "sentence_length_variance": 5.2,
    "avg_message_length_words": 21,
    "formality_score": 0.32,
    "emoji_frequency": 0.18,
    "top_emojis": ["👍", "🔥", "😊"],
    "punctuation_style": {
      "period": 0.45,
      "exclamation": 0.31,
      "question": 0.19,
      "ellipsis": 0.05
    },
    "common_openers": ["hey", "hi", "yo", "quick question"],
    "common_closers": ["thanks", "lmk", "let me know", "cheers"],
    "contraction_rate": 0.67,
    "vocabulary_uniqueness": 0.41
  },
  "sample_count": 487,
  "last_trained_at": "2026-02-14T15:35:12Z"
}

How to interpret these numbers:

formality_score: 0.0 = very casual, 1.0 = very formal. Most business users land 0.3-0.7.
emoji_frequency: 0.18 = ~18% of messages contain emojis (higher than average)
contraction_rate: 0.67 = you use "don't" instead of "do not" about 2/3 of the time
vocabulary_uniqueness: How often you use uncommon words (0.41 is moderate—some jargon, mostly plain language)

Red flags:

avg_sentence_length over 20: Your messages are too verbose, AI will ramble
formality_score mismatch: If you're casual but it reads formal (0.8+), training data had too many templates
Low sample_count (under 50): Retrain with more data

If the profile looks wrong, delete it and retrain with filtered data.

Before and After: Real Examples

Customer question: "Do you offer refunds?"

Generic AI (no Learn Mode): "Thank you for your inquiry regarding our refund policy. Yes, we do offer refunds within 30 days of purchase for unused services. Please contact our support team to initiate the refund process, and they will assist you with the necessary steps."

With Learn Mode (casual style profile): "yep! 30 days for refunds if you haven't used it. just shoot us a message and we'll sort it out 👍"

With Learn Mode (formal style profile): "Yes, we have a 30-day refund policy for unused services. Please reach out to our support team to process your refund. Happy to help."

Same factual content. Three completely different vibes. Learn Mode adapts to match your natural writing patterns.

Step 4: Deploy and Monitor

Enable Learn Mode for your AI auto-replies by setting apply_style: true:

bash

curl -X POST https://apiv2.waiflow.app/api/v2/ai/generate-reply \
  -H "Authorization: Bearer YOUR_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "contact_id": "[email protected]",
    "context": "Do you have pricing for teams?",
    "use_rag": true,
    "apply_style": true
  }'

Response includes style confirmation:

json

{
  "reply": "yeah we do team pricing! starts at $89/mo for up to 5 people. want me to send you the full breakdown?",
  "rag_sources": ["pricing-sheet.pdf"],
  "style_applied": true,
  "model": "gpt-4o-mini",
  "tokens_used": 203
}

Notice style_applied: true—this confirms Learn Mode was active for this response.

Monitor quality over time:

Check your AI responses weekly to ensure Learn Mode isn't drifting:

Are responses still matching your voice?
Any awkward phrasings?
Customer feedback (do they realize it's AI?)

When to retrain:

Your writing style evolves (you get more casual/formal)
You add 100+ new messages (fresh data improves accuracy)
Customer feedback indicates AI sounds "off"
You change business context (new product, new audience)

I retrain my style profile every 3 months or whenever I notice quality degradation.

Common Issues and Fixes

Issue: AI responses are too casual for customers

Your personal style might be too informal for customer-facing use. Solutions:

Train a separate profile on customer support messages only
Use contact-specific profiles (casual for internal, formal for customers)
Adjust prompts to override style (prompt takes precedence)

Issue: AI isn't using emojis even though I do

Check emoji_frequency in your profile. If it's low (under 0.10), the training data didn't have enough emoji examples. Options:

Use more emojis in your daily messages (seriously)
Manually add emoji preferences to your prompt
Accept that AI will be slightly more conservative with emojis

Issue: Responses are too long

Your avg_message_length_words is probably high. Two fixes:

Train on shorter messages only (filter out long explanations)
Add length constraints to your prompt: "Keep responses under 30 words when possible"

Issue: Learn Mode makes AI responses inconsistent

This happens when your training data has high variance (you write very differently depending on context). Solution:

Create multiple style profiles for different contexts
Use contact-specific training
Disable Learn Mode for contexts where consistency matters more than personality

When to Use Learn Mode vs Custom Prompts

Learn Mode and custom prompts solve different problems. Here's when to use each:

Use Learn Mode when:

You want AI to sound like you personally
Maintaining your brand voice matters
You're a solopreneur or small team with consistent voice
You write to customers directly (not through templates)

Use custom prompts when:

You need specific factual accuracy
Compliance requires exact phrasing
Multiple people use the account (no single "voice")
You're selling complex products (need detailed explanations)

Use both when:

You want your voice but need to enforce business rules
Prompt defines what to say, Learn Mode defines how to say it
Most common setup for Pro/Business users

Example combining both:

Prompt: "You're a sales assistant for Acme Corp. Always mention our 30-day guarantee when discussing pricing."

Learn Mode: Ensures the mention sounds natural, not scripted.

Result: "yeah our pro plan is $99/mo, and btw you get a 30-day guarantee so zero risk 👍"

The prompt ensured the guarantee was mentioned. Learn Mode made it sound human.

What's Next

Learn Mode works great standalone, but it's even better combined with other MoltFlow features:

Enhance your AI:

Set up AI Auto-Replies — Combine Learn Mode with intelligent responses for 24/7 support
Build a Knowledge Base AI — Ground responses in your documentation while keeping your voice
Train AI Writing Style — Step-by-step guide to optimizing Learn Mode

Scale automation:

AI Model Comparison — Choose between GPT-4, Claude, and Gemini for quality vs cost
WhatsApp 2026 Compliance Guide — Keep your bot compliant with Meta's policies

Ready to make your AI sound human? Upgrade to Pro — Get Learn Mode, RAG, and unlimited messages. 14-day free trial, no credit card required.

Key Takeaways

Learn Mode isn't a replacement for good prompts—it's a multiplier. You still need to define what the AI should say and when to escalate. Learn Mode just makes it sound like you said it.

Start with 200+ quality messages, train a global profile, review the features to ensure accuracy, then deploy with monitoring. Retrain quarterly or when you notice drift.

The difference between generic AI and AI trained on your voice? Customers stop asking "am I talking to a bot?" That's worth the 5-minute setup.

Questions? Hit us up in Discord or check the API docs.