Learn the author's natural writing voice from their past papers and apply it as a soft guide when drafting new text. Use when drafting or polishing paper sections and the user wants the text to sound like their own writing, or when a Style Profile should be built/refreshed from writing samples. Personalization, not de-AI-ification.
Installs into .claude/skills of the current project.
Are you the author of Style Calibration?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/yichen-zju-style-calibration)
---
name: style-calibration
description: Learn the author's natural writing voice from their past papers and apply it as a soft guide when drafting new text. Use when drafting or polishing paper sections and the user wants the text to sound like their own writing, or when a Style Profile should be built/refreshed from writing samples. Personalization, not de-AI-ification.
argument-hint: [path-to-writing-samples]
---
# Style Calibration(作者声纹校准)
从作者过往写作样本学习自然声纹,起草时作为软性指导。完整协议见 `references/style_calibration_protocol.md`(改编自 ARS, CC-BY-NC 4.0)。
**设计边界**:这不是 AI 降痕/洗稿工具。目标是让文本像作者本人写的——作者的判断与风格是学术身份的一部分。学科规范优先于个人习惯。
## 使用方式
1. **建档**(一次性,可复用):收集作者 2-5 篇过往论文/草稿(用户提供路径),提取 Style Profile:
- 句长分布与节奏(长短句配比)
- 段落开头习惯(直入主题 vs 背景铺垫)
- 术语偏好与惯用搭配(如 "we propose" vs "this paper presents")
- 论证习惯(先 claim 后证据 vs 先铺垫后 claim)
- 图表引用风格、hedging 习惯("may" / "suggests" 的使用频率)
- 明确的"从不这样写"清单
2. **保存**:`outputs/style-profile-<author>.md`,后续项目复用。
3. **应用**:起草/润色时把 Profile 作为软约束——`paper-production` Stage 4 在调用 writer 前读取它(如果存在)。
4. **校准**:新写的段落回头和样本对比,更新 Profile。
## 纪律
- 样本必须是作者本人的作品;引用他人风格需用户明示。
- Profile 是软指导:与证据严谨性、学科规范冲突时让路。
- 保留作者修订痕迹:作者改掉 AI 文本的地方是最好的新样本。