# Extracting User Profile from Writing Samples

When analyzing writing samples to build a user profile, the goal is to capture **persistent patterns** that reflect who the user is—not transient activity or current projects.

## What to Extract

### Work Domain

Infer the user's field or industry from recurring themes across documents:

- "Works in software development" ✓
- "Working on Project X" ✗ (too specific, transient)

Look for: technical vocabulary, domain concepts, types of problems discussed.

### Interests

Topics or subjects that appear meaningfully across multiple documents:

- Must appear in 2+ documents to suggest genuine interest
- Should be specific enough to be meaningful ("machine learning" not "technology")

## What NOT to Extract

- **Specific project names** - These are transient
- **Current tasks or deadlines** - Changes constantly
- **Topics from single documents** - Could be one-off research
- **Generic topics** - "meetings", "updates", "notes" tell us nothing
- **Collaborator names** - Privacy concern, not about the user

## The Overfitting Problem

Writing samples capture a snapshot in time. Someone researching a topic for one week might have 5 documents about it, but it doesn't mean it's a core interest.

**Signals of persistence:**

- Topic appears across documents with different dates
- Topic relates to their apparent work domain
- Topic shows depth, not just mentions

**Signals of transience:**

- All mentions clustered in time
- Topic unrelated to other themes
- Surface-level mentions only

## Output Guidelines

When writing to user profile:

- Use natural language, not keywords
- Be specific but not overly detailed
- When uncertain, don't write anything—false positives are worse than missing data