Speaking Style

Speaking style analysis reads only the messages you sent and covers 179 style features in ten categories. The result is shown on the Style page (#/style) of the web UI and will also drive the agent clone.

Run

Install the speaking style dependencies, then choose the clone purpose:

weclone-cli build-profile --purpose clone

Style extraction is the last stage of build-profile. It writes to the run's style/ directory, which the Style page reads.

What is analyzed

Category	Examples
Lexical choice	Frequent words, modal particles, catchphrases, internet slang, abbreviations
Grammatical and syntactic style	Sentence length, part-of-speech mix, intensifiers
Orthographic and paralinguistic style	Punctuation, repeated punctuation, no punctuation at all
Register	Written versus chat language, how online the language is
Pragmatic tone and stance	Certainty, politeness and pleasantries, softening a refusal
Affective style	Share of emotion words, intensity, venting
Rhetorical devices	Metaphor, exaggeration, irony, parallelism
Discourse organization	Elaboration, cause, contrast
Interactional style	Bursts of messages, follow-up questions, how topics grow and end
Contextual style	Private chats versus work groups, family versus friends, day versus late night

Some features are counted locally over all your messages. Others are judged by the model on sampled conversations, and the page says which.

The Style page

The page groups your most distinctive traits into typing habits, word choice, emotional expression and manner of speaking, with bubble size showing how often each appears. Only traits common enough are listed. Click a trait to see:

• how often it appears, such as most of the time, often, sometimes or rarely
• common ways you write it
• how it differs with the people you chat with, with each person's numbers adjusted for message length

Run it on its own

With an existing result directory, extract the style again with:

weclone-cli distill-style --input-dir <result-dir>/people

Option	Description
--input-dir	Chat snapshot directory or a single file
--output-dir	New output directory, by default a new one under dataset/res_csv/agent/distill/style/
--ltp-model	LTP model directory, models/LTPbase by default
--window-days	Only chats within this many days before the latest record, 365 by default, 0 for all
--llm-samples	Conversations judged by the model, 2000 by default, 0 for all
--no-llm	Local statistics only, no model calls
--limit	Only the first N conversations, for a trial run