My AI Avatar: Trained from 35,000+ WeChat Messages

Start: can data tell me who I am?

We all claim to "know ourselves", but when asked, we just throw adjectives. I wanted a different route: export everything I've said, and let data answer "who am I".

The result is the AI avatar you can chat with at the bottom-right of this site — its speaking style comes from this training.

Step 1: getting the records out (the painful part)

WeChat stores chats locally but encrypted. Two hurdles:

Tool selection. The ecosystem is turbulent: old tools stopped, WeFlow got DMCA'd by Tencent, PyWxDump abandoned. The only active one: CipherTalk (密语) — with a CLI for scripted export.

Getting the key. The decryption key exists only at the instant of login. You must: log out → stay on QR screen → arm the key hook → scan & log in → catch the key in that instant.

Step 2: let the data speak

I wrote Python to analyze my own 35,659 messages (21,335 after filtering system text):

• Avg length — 22.3 chars (36% under 8)
• Favorite particles — 吧 / 哈哈 / 我觉得 / 确实 / 可以的
• Emoji use — 2% · ❤ 👌 ✅ 😋
• Active hours — afternoons (40%)

These stats are more honest than my self-description: I do speak short, use particles, and occasionally emoji.

Step 3: make it reusable

I packaged this into three layers: style profile (readable stats), my-* skills (write replies in my voice, portrait generator), and an Obsidian vault of chat notes.

Step 4: put the avatar to work

The site's chat avatar distills this style into a system prompt: short, colloquial, with particles, slightly restrained for the job-search context. It answers "what projects have you done?" the way I would — because it learned how I talk.

Two lessons

1. Data is the self. What we think we know often loses to what statistics reveal.
2. Privacy is a line. This pipeline studies abstract style (tone, length, habits) — not specific conversations. The avatar knows how I speak, not who I talked to. Learn the style, never expose the content.