Your writing style is as unique as your signature. But how can an algorithm learn it automatically?
When Neston starts, it scans your 300 sent and 500 received emails. In 3 minutes, Claude analyzes your writing patterns: your average sentence length, preferred vocabulary, punctuation style, formality level, even your favorite expressions.
The result: every generated email sounds like you — not like ChatGPT.
📖 Table of Contents
Machine Learning Principles Applied to Emails
Machine learning is the art of teaching an algorithm to discover patterns without explicitly writing every rule. You don't say "if a sentence is more than 25 words, it's too long" — you let the algorithm discover the patterns in your data.
At Neston, we use a 3-phase approach:
Phase 1: Initial Scan (Onboarding)
On first launch, Claude scans your last 800 emails (300 sent + 500 received). This 800-email window represents your "writing signature" — enough data to see a pattern, without being polluted by old habits you've changed.
Phase 2: Profiling
The AI creates a style profile with 5 dimensions:
- Formality: Do you use "Sincerely" or "Cheers"?
- Length: Are your emails 100 words or 500 words?
- Vocabulary: Technical, commercial, educational?
- Tone: Direct, empathetic, humorous?
- Structure: Numbered lists? Bullets? Long paragraphs?
Phase 3: Personalized Generation
When you request a reply, Claude uses this profile as a generation constraint. You get text that respects your natural writing patterns.
Tokenization and Lexical Analysis
Before even "understanding" style, the AI must first break text into pieces. This process is called tokenization.
This step seems trivial, but it's crucial. The AI then counts:
- How often do you use informal greetings ("Hi") vs. formal ones ("Dear")?
- Are your sentences short (8 words average) or long (25 words)?
- Do you use lots of punctuation or minimal?
- Do you have writing "tics"? ("Actually", "Obviously", "Exactly")
These lexical statistics form the foundation of your writing profile.
Embeddings: Writing Vectors
An "embedding" is a mathematical representation of what your style is. Imagine each email as a point in 768-dimensional space (for Claude).
Your writing space looks like this:
- A very formal email = position A
- A very informal email = position B
- An average email = position between A and B
When Claude generates an email, it "places" the generated text in the right spot in your space — not too formal, not too informal, just you.
Analogy: If your style is "a color," Claude learns where on the red-blue-green spectrum you sit. Then, every generated sentence is tinted exactly in your color.
Continuous Learning Cycle
The initial profile is good, but insufficient. That's why Neston learns continuously.
After Each Email Sent
When you approve and send a generated email, Neston records:
- What did you change? (Edits = improvement signal)
- How did Claude misunderstand your style?
- This correction improves the model for your next emails
After 10 Emails
Every 10 emails, the system recalculates your profile with new data. Your style 3.0 = style 2.0 + learning from the last 10 sends.
After 50 Emails (Major Calibration)
After 50 emails, complete recalibration. The system reanalyzes your correction habits: which sentence types do you modify? How do you correct them? This analysis creates a powerful feedback loop.
Result: after 50 emails, acceptance-without-modification rate jumps from 60% to 85%+
* Estimate based on our internal model — results vary with usage.
Progressive Adaptation
Key point: learning is not linear.
- Your first 10 emails teach 40% of your total style
- The next 10 teach 30% more
- The next 30 teach the remaining 30% (nuances, exceptions) * Indicative distribution based on our internal learning model.
It's like learning a language: you master 80% of vocabulary with 1000 words, but the remaining 20% takes 10,000 words.
💡 Key Stat: After 50 processed emails, Neston's AI model knows you better than 99% of generic tools. After 100, it's nearly impossible to distinguish between a generated email and a real one from you.
Limits of Machine Learning
No illusions: machine learning has real limits.
1. Data Contamination
If you have 300 sent emails but 250 are forwarded meeting summaries, the profile is polluted. Neston applies quality filters to identify and exclude this noise.
2. Natural Evolution
You changed jobs and your style transformed? The model adapts progressively, but there's a lag. No "complete retroactive learning."
3. Unusual Emails
If you suddenly write something totally new (ex: an email in English when you usually write French), the AI might struggle. That's normal — you're making an exception.
In Summary: Understanding Your AI Learning
- Tokenization: AI breaks your emails into words and sentences to analyze patterns
- Embeddings: Your style is converted into mathematical vectors (your "writing space")
- Continuous cycle: After 10, 50, 100 emails, the profile becomes increasingly precise
- Final result: Emails that sound 100% you, not 100% ChatGPT
Curious to see how Neston learns your style in action?
Launch a free trial. After the 1st email, you'll see the difference: Claude already knows your patterns.
Discover AI Learning →14 days free · Machine learning in action
Your writing style isn't a bug to fix — it's your unique value. Neston preserves and amplifies it.