逆向工程你的思维、言语和决策——然后将其转化为经过压力测试的AI系统提示词。适用于任何大型语言模型。完整流水线，包含Claude代码。
概览
*本片中的每一个数字和句子都是该工具虚构的样本写作输出。由tools/demo/make_demo.py重建。* AI负责判断。所有能计数或检查的都用代码完成：它运行在你的机器上，每次给出相同的答案，并在保存任何东西前就剔除邮件、电话号码和金额。 你需要 Python 3.9 或更新版本。没有任何东西需要安装。没有任何东西会离开你的机器。 第三步故意失败：草稿会对抗，使用破折号，然后用一句话写成37个单词。examples/drafts/supplier-update.fixed.md 通过了。 要在自己的写作中运行，复制你自己的.md或.txt文件到一个文件夹，清理它们（见安全），然后指向那个文件夹——-corpus。 提取本身就是一个提示。深度取决于模型能读取什么。 1. 把EXTRACTION_PROMPT.md复制到新聊天里。2. 把你的文字粘贴到底部。五页或更多。你平时的写作，不是最好的。3.
概述（AI 翻译）
看看你实际写作的方式。然后对照你的或AI的每一个草稿。
本片中的每一个数字和句子都是该工具虚构的样本写作输出。由tools/demo/make_demo.py重建。
你得到了什么
| 你得到 | 是什么 |
|---|---|
| 关于你写作的报告 | 一页，任何浏览器都能打开。你的句子长度，你多频繁地保留，你依赖哪些短语，哪些文档听起来与其他文档不同，以及需要修正的具体行数。 |
| 一个任何AI都能使用的双胞胎 | 一个系统提示，捕捉你的声音和决策方式，基于你的写作和最多五个关于写作未显示内容的问题构建。 |
| 每个草稿都有检查器 | 规则代码可以逐行检查。可以在草稿中运行，或者让它检查每个拉取请求，这样就能在发货前抓到语音外的副本。 |
| 一个跟上节奏的双胞胎 | 在你实际发送的内容旁边显示草稿。它会根据你的编辑提出新规则。你批准每一条规则，所有更改都会被记录。 |
AI做判断。
本页由 AI 翻译，技术细节请以原文为准。
README
以下为原文 README 节选，完整内容请查看原文。
Digital Twin of Yourself
See how you actually write. Then check every draft, yours or an AI’s, against it.
Every number and sentence in this film is the tool’s real output on made-up sample writing. Rebuilt by tools/demo/make_demo.py.
What you get
| You get | What it is |
|---|---|
| A report on your writing | One page, opens in any browser. How long your sentences run, how often you hedge, which phrases you lean on, which documents sound unlike the rest, and the exact lines to fix. |
| A twin any AI can use | A System Prompt that captures your voice and how you decide, built from your writing plus up to five questions about what the writing does not show. |
| A checker for every draft | Rules code can check, line by line. Run it on a draft, or let it check every pull request so off-voice copy is caught before it ships. |
| A twin that keeps up | Show it drafts next to what you actually sent. It proposes new rules from your edits. You approve each one, and every change is logged. |
The AI does the judgment calls. Everything that can be counted or checked is done by code: it runs on your machine, gives the same answer every time, and strips emails, phone numbers and money amounts before anything is saved.
Built by WhyStrohm.
Try it in three minutes
You need Python 3.9 or newer. Nothing to install. Nothing leaves your machine.
git clone https://github.com/whystrohm/digital-twin-of-yourself.git
cd digital-twin-of-yourself
# 1. Measure a folder of writing (here, the made-up sample)
python3 scripts/twin_scan.py --corpus examples/sample-corpus --out patterns.json
# 2. Build the report and open it
python3 scripts/twin_report.py patterns.json --out twin-report.html
open twin-report.html # macOS; use xdg-open on Linux
# 3. Check a draft against a twin's rules
python3 scripts/twin_check.py --rules twins/example/twin.rules.json examples/drafts/supplier-update.md
Step 3 fails on purpose: the draft hedges, uses a dash, and runs one sentence to 37 words. examples/drafts/supplier-update.fixed.md passes.
To run it on your own writing, copy your own .md or .txt files into one folder, scrub them (see Safety), and point --corpus at that folder.
The layers
The extraction itself is a prompt. How deep it goes depends on what the model can read.
| Layer | Where | What it reads | Who measures |
|---|---|---|---|
| 1 | Any LLM (ChatGPT, Gemini, Claude) | Writing you paste in | The model estimates |
| 2 | Claude with memory on | Your conversation history | The model estimates |
推荐工具
换一个关键词，或者移除筛选条件。
安装
npx skillfish add whystrohm/digital-twin-of-yourself