chat-distill（聊天提炼）

Chat Distill — Style Analysis & Mimicry

Workflow

1. Parse → extract messages per speaker from raw export (see references/format-parsers.md)
2. Analyze → build style profile (see references/style-dimensions.md)
3. Report → output analysis report using template in references/output-template.md
4. Mimic → generate replies on demand using the profile

Quick Start

Given a chat export file:

1. Read the file and identify the format (WeChat export, plain text, JSON array, TG export).
2. Normalize into { speaker, text, time? } messages using parsing rules in references/format-parsers.md.
3. Pick the target speaker — the one whose style to learn. If multiple speakers exist, ask which one.
4. Run analysis following references/style-dimensions.md.
5. Output the report per references/output-template.md § Analysis Report.
6. When the user asks for a mimicked reply, use the profile + references/output-template.md § Mimic Reply.

Key Principles

• Show, don't tell: Include concrete examples from the actual chat when reporting style traits.
• Preserve quirks: Capture tics the speaker doesn't notice — repeated filler words, capitalization habits, punctuation style.
• Respect privacy: Never echo sensitive content (passwords, addresses, financials) from chats into reports. Anonymize if needed.
• Minimum sample: Require at least 20 messages from the target speaker. If fewer, warn that analysis may be unreliable.