per-person activity
message volume, words written and average length per person
| person | messages | words | w/day | avg/msg | media | questions | median reply |
|---|
10 years of messages between 3 friends
how this was made
Built with Claude Code over several iterations. The pipeline today:
- Parse raw WhatsApp
.txtexports into a per-message data blob; sender names aliased. - Claude Sonnet reads the chat in 100-message windows to find conversation boundaries — about 1,860 conversations across ten years.
- Stratify-sample 200 conversations, have Sonnet draft a topic taxonomy from the sample, hand-edit it once, then lock it (12 categories × 75 subtopics).
- Tag each conversation against the locked taxonomy via roughly 3,000 prompt-cached Sonnet calls.
- Short reactions ("lol", "yeah", "haha") inherit their conversation's topic instead of being tagged in isolation.
word share
who dominates the conversation, weighted by words
who starts chats
first message of a new calendar day counts as a fresh start
activity over time
share per person (bars, left axis) · total messages (dotted line, right axis), monthly
hour of day
when this chat is most alive
day of week
mon → sun
topics
LLM-classified categories — word-share per person. messages may carry multiple tags.