Paper Bilingual

将PDF论文转换为双语Markdown。对用户提供PDF URL或本地PDF文件，提取文本转为Markdown，保留图片，逐段插入中文翻译。用于学术论文研读、翻译对比。

Security Scan

The skill's requirements and instructions are internally consistent with its stated purpose of converting PDFs to bilingual Markdown; it requires local file writes and Python libraries but asks for no unrelated credentials or installs.

Purpose & Capability

The name/description (PDF -> bilingual Markdown, preserve images) matches the SKILL.md workflow: download (if URL), extract text and images with PyMuPDF, crop figures, then translate paragraphs with the current LLM and write Markdown. All requested operations are expected for this use case.

Instruction Scope

Instructions are specific and limited to the task (downloading the supplied URL, parsing PDF, saving images and Markdown, translating per-paragraph). Minor scope notes: the download example uses PowerShell (Windows-specific) while the skill declares no OS restriction; translation is done via the current session LLM which may hit token limits for large PDFs.

Install Mechanism

This is instruction-only with no install spec (lowest install risk). It lists dependencies (Python 3.8+, PyMuPDF, Pillow) but does not provide an automated install step; users will need to pip-install them themselves.

Credentials

No environment variables, credentials, or config paths are requested. The only external network action is downloading the user-provided PDF URL, which is consistent with the stated purpose.

Persistence & Privilege

always is false and the skill does not request elevated or persistent privileges. It writes files under a per-paper folder (memory/paper-bilingual/...), which is reasonable for its function.

Assessment

This skill appears coherent for converting PDFs to bilingual Markdown. Before using it: ensure you trust the PDF source (the skill will download the provided URL), have Python 3.8+ and install pymupdf and pillow with pip, and be aware it will write files under memory/paper-bilingual/ on your filesystem. The download example uses PowerShell (Windows); adapt the download step on macOS/Linux. For very long papers, per-paragraph translation via the LLM may hit token or cost limits — consider splitting work or translating selectively. No credentials or hidden network endpoints are requested by the skill.

将PDF论文转换为中英双语Markdown，保留原文格式和图片，逐段提供中文翻译。

触发条件

用户要求以下操作时使用此skill：

• "翻译这篇论文"
• "把PDF转成双语MD"
• "论文中英对照"
• "下载论文转Markdown带翻译"
• 提供PDF URL并要求翻译

输入

支持两种输入方式：

1. URL输入：PDF的下载链接（http/https开头）
2. 本地路径：本地PDF文件的绝对或相对路径

处理流程

步骤1：下载PDF

• 如果是URL，使用PowerShell下载：

Invoke-WebRequest -Uri "URL" -OutFile "$env:TEMP\paper.pdf"

• 如果是本地路径，直接使用

步骤2：PDF转Markdown

使用Python + PyMuPDF将PDF转为Markdown并提取图片：

import fitz
import os
from PIL import Image

pdf_path = "paper.pdf"
images_dir = "images"
figures_dir = "figures"
os.makedirs(images_dir, exist_ok=True)
os.makedirs(figures_dir, exist_ok=True)

doc = fitz.open(pdf_path)
pages = []

for page_num, page in enumerate(doc):
    # 提取文本
    text = page.get_text("text")
    
    # 提取图片
    images = []
    for img_index, img in enumerate(page.get_images()):
        xref = img[0]
        pix = fitz.Pixmap(doc, xref)
        if pix.n - pix.alpha > 0:  # 跳过透明图片
            img_name = f"page{page_num + 1}_img{img_index}.png"
            pix.save(os.path.join(images_dir, img_name))
            images.append(img_name)
        pix = None
    
    pages.append({'text': text, 'images': images})

步骤3：提取完整图表（推荐方法）

使用整页渲染+坐标裁剪提取完整图表：

import fitz
from PIL import Image

pdf_path = "paper.pdf"
OUTPUT_DIR = "figures"
os.makedirs(OUTPUT_DIR, exist_ok=True)

doc = fitz.open(pdf_path)
fig_count = 0

for page_num in range(len(doc)):
    page = doc[page_num]
    # 2x缩放渲染页面为高清图片
    pix = page.get_pixmap(matrix=fitz.Matrix(2, 2))
    img = Image.frombytes('RGB', [pix.width, pix.height], pix.samples)
    
    # 获取页面上的图片对象和位置
    images = page.get_images()
    
    for img_index, img_info in enumerate(images):
        xref = img_info[0]
        rects = page.get_image_rects(xref)
        for rect in rects:
            # 缩放坐标 (2x)
            x0 = max(0, int(rect.x0 * 2) - 20)
            y0 = max(0, int(rect.y0 * 2) - 20)
            x1 = min(pix.width, int(rect.x1 * 2) + 20)
            y1 = min(pix.height, int(rect.y1 * 2) + 20)
            
            # 过滤太小的图片碎片
            if x1 - x0 < 80 or y1 - y0 < 80:
                continue
            
            # 裁剪完整图表
            cropped = img.crop((x0, y0, x1, y1))
            fig_name = f'fig_p{page_num+1}_{img_index+1}.png'
            cropped.save(os.path.join(OUTPUT_DIR, fig_name))
            fig_count += 1

print(f"Extracted {fig_count} figures")

关键点：

• PDF里的图表可能被拆成多个小图片块（tiles）压缩
• 直接提取会得到碎片，需要用坐标裁剪方法
• 先渲染整页，再按图片坐标裁剪，确保图表完整

步骤4：翻译并生成双语MD

输出目录结构：

memory/paper-bilingual/
└── YYYY-MM-DD-paper-name/
    ├── index.md          # 双语Markdown
    ├── images/           # 整页截图（可选）
    └── figures/          # 裁剪后的图表

输出格式：

# 主标题
中文主标题

## Section 1
中文标题

原文段落...
中文翻译段落...

![Figure 1](figures/fig_p3_1.png)

原文段落...
中文翻译段落...

规则：

1. 主标题（#）：英文后换行加中文
2. 二级标题（##）：英文后换行加中文
3. 普通段落：原文后直接加翻译
4. 图片：使用裁剪后的图表 figures/

步骤5：翻译实现

使用当前会话的LLM翻译，逐段调用：

将以下英文翻译成中文，保持自然段落格式：
[段落内容]

步骤6：保存输出

• 根据PDF文件名或URL生成输出目录名
• 保存到：memory/paper-bilingual/YYYY-MM-DD-[论文名]/
• 清理临时文件

核心原则

1. 每篇论文一个文件夹：避免不同论文文件冲突
2. 原文完全保留：图片、代码、公式、表格等与源文件完全一致
3. 逐段翻译：每段正文后插入该段的中文翻译
4. 标题翻译：所有级别的标题都需要翻译
5. 格式一致：翻译采用与原文相同的格式
6. 完整图表：使用坐标裁剪方法提取完整图表，避免碎片

依赖

• Python 3.8+
• PyMuPDF (pip install pymupdf)
• Pillow (pip install pillow)
• LLM翻译（当前会话）

错误处理

• URL无效：提示用户检查链接
• PDF加密：提示需要解密
• 翻译失败：跳过该段，保留原文