A skill set that helps agents become your AI twin🪞

中文说明

WeClone-Skill uses a lightweight persona pack to turn "reply like a specific person" into a maintainable, reviewable workflow. Define the persona and boundaries first, inject the live reply context second, and keep a human review step at the end. The repository centers on two skills:

• 🧩 weclone-init-twin: initialize a persona pack template
• ✍️ weclone-twin-reply: combine persona files with the current scene into an isolated prompt for reply drafting

✨ Use Cases

• 💬 Companion-style chats for close friends or family on WhatsApp / QQ and similar apps
• 🌐 Creator or social DM replies that preserve a familiar tone when thanking or encouraging followers
• 📧 Workplace email drafting and review, including current constraints and next steps
• 🤝 Sales follow-up and quoting workflows across HubSpot leads or WeCom conversations
• 🕊️ Digital legacy persona preservation for memorial and companionship-style conversations based on a deceased person's historical data

Installation

Quick install (recommended)

npx skills add xming521/WeClone-Skills

Or tell your agent directly

Please help me install the skills from github.com/xming521/WeClone-Skills

Workflow

1. Initialize a persona pack

Use weclone-init-twin to generate a set of markdown files:

• profile.md: identity, personality, conversation style, values, worldview, decision style
• persona_examples.md: real reply examples from the person being cloned
• state.md: recent status, current goals, short-term constraints
• guardrails.md: hard boundaries, promises the twin must not make, privacy limits

2. Fill the pack with real content

The template is only the skeleton. Reply quality depends heavily on whether these files are concrete, real, and information-dense.

3. Add runtime context

Before each generation, add two kinds of runtime information:

• scene.md: a concise description of the current scene and background
• dialogue.md: the active conversation window that needs a reply

4. Render an isolated prompt

Use weclone-twin-reply to combine persona files and runtime context into a single prompt package for a downstream model to draft from in isolation.

5. Review before sending

This repository keeps the review gate by default. It drafts replies; it does not silently send them.

Quick Start

Initialize a persona pack

Generate the default English ai_twin/ directory from the repository root:

python3 skills/weclone-init-twin/scripts/init_twin_profile.py --user-name "Alex"

Generate a Chinese persona pack in a custom directory:

python3 skills/weclone-init-twin/scripts/init_twin_profile.py \
  --user-name "张三" \
  --language zh \
  ./my_twin

If the target files already exist, the script refuses to overwrite them unless you pass --force.

Render a prompt from a persona pack

python3 skills/weclone-twin-reply/scripts/render_clone_prompt.py \
  --persona-dir ./ai_twin \
  --scene ./scene.md \
  --dialogue ./dialogue.md

Write the rendered prompt to a file:

python3 skills/weclone-twin-reply/scripts/render_clone_prompt.py \
  --persona-dir ./ai_twin \
  --scene ./scene.md \
  --dialogue ./dialogue.md \
  --output ./rendered_prompt.md

Add extra context only when it will materially affect the reply:

python3 skills/weclone-twin-reply/scripts/render_clone_prompt.py \
  --persona-dir ./ai_twin \
  --scene ./scene.md \
  --dialogue ./dialogue.md \
  --extra-context ./constraints.md

Included Examples

The repository includes Chinese and English examples. See examples/README.md:

• examples/zh/social-content-twitter-fan-dm
• examples/zh/workplace-email-draft
• examples/zh/sales-copy-hubspot-wecom
• examples/zh/companion-chat-whatsapp-qq
• examples/zh/digital-legacy-memorial-reply
• examples/en/social-content-twitter-fan-dm
• examples/en/workplace-email-draft

Example command:

python3 skills/weclone-twin-reply/scripts/render_clone_prompt.py \
  --persona-dir examples/zh/social-content-twitter-fan-dm \
  --scene examples/zh/social-content-twitter-fan-dm/scene.md \
  --dialogue examples/zh/social-content-twitter-fan-dm/dialogue.md

Design Principles

• Separate long-term persona information from live conversation context
• Use guardrails.md to make behavioral boundaries explicit
• Keep runtime context short, but high-signal
• Preserve human review before anything is sent externally