BellFlower-Skills

Convert Markdown with LaTeX math, images, and tables into a polished PDF. Two rendering engines: (1) weasyprint — pure Python, no browser needed; (2) chromium — best for matrices/bmatrix/pmatrix/vmatrix and complex formulas. Supports Chinese/Japanese fonts. Can chain after ocr-md-polish for a complete OCR→clean→PDF workflow.

Burn core framework - Provides tensor operations, automatic differentiation, and neural network building blocks for Rust deep learning.

Burn CUDA backend - High-performance NVIDIA GPU acceleration. Provides optimal performance for NVIDIA hardware.

Burn Metal backend - Apple GPU acceleration for macOS and iOS. Optimized for Apple Silicon and Intel-based Macs.

Burn NdArray backend - CPU-based tensor operations using Rust's ndarray crate. Ideal for development, testing, and lightweight inference.

Burn ONNX support - Import ONNX models into Burn for inference on any backend. Supports compile-time code generation and runtime loading.

Provides guidance on using Burn deep learning framework for Rust projects. Invoke when working with tensor operations, model inference, or ONNX model integration in Rust.

Burn WGPU backend - Cross-platform GPU acceleration using WebGPU. Supports Vulkan, Metal, Direct3D, and WebGL.

[DRAFT] Python→Rust 渐进式重写总 skill。定义核心思想、链路、运行时模型、上下文管理机制与状态机；统一调度依赖对齐、跑通环境、运行时、7轮审查等子 skill。

[DRAFT] 对齐 Python / Rust / FFI / async / 错误追踪 三方依赖；生成清单；优先用 cargo add 锁定最新版。py2rs 流水线 Stage 0 的前半部分。

[DRAFT] 打通 Python/Rust 混合运行环境。用最小 Demo 验证 FFI 调用、参数/返回值/错误传递、async 调度、日志追踪。py2rs 流水线 Stage 0 的后半部分。

[DRAFT] 第 0 轮审查：行为一致性验证。证明 Python 实现 == 迁移版 Rust 实现。不允许改代码，仅产出对比报告。

[DRAFT] 第 1 轮审查：把 Python 风格的 Rust 改写为工程化 Rust。引入 mod.rs / lib.rs / trait / crate 划分 / 生命周期整理。允许结构层面小改，不允许改业务语义。

[DRAFT] 第 2 轮审查：统一错误与追踪体系。引入 tracing/anyhow/thiserror，让错误能定位到文件 / 函数 / 调用链。允许小改。

[DRAFT] 第 3 轮审查：消除阻塞 IO 与不必要等待。引入 async、tokio、rayon。允许 IO 层面大改；允许算法逻辑小改以匹配 IO。

[DRAFT] 第 4 轮审查：算法复杂度与效率（计算机科学家视角）。只在有复杂度分析 + 理论依据 + Benchmark 三件套时才允许修改算法。

[DRAFT] 第 5 轮审查：架构与数据结构（资深软件工程师视角）。允许数据结构大改；关注所有权模型、Arc<Mutex> 使用、API 设计。

[DRAFT] 第 6 轮审查：产品与人体工学。从用户视角审视批处理、缓存、国际化、自动恢复、UI/CLI 体验。只审查和分析，不写代码，输出报告给用户。

[DRAFT] 建立 py2rs 的运行时骨架：manifest 注册、迁移状态机读写、路由协议、跨语言调用桥。是 py2rs 总 skill 的唯一事实来源和决策层。

Polish an OCR-generated Markdown file that already has LaTeX math. Fix formula rendering issues (subscripts, grouping, nested delimiters), remove OCR-duplicated text near images, clean up broken tables, restore list numbering hierarchy, unify split inline math, validate formulas semantically against surrounding context.

Translate Markdown text (lecture notes, documentation, articles) between languages while preserving Markdown formatting, LaTeX math, code blocks, images, HTML comments, and frontmatter. Use this skill whenever the user asks to translate a Markdown file, convert lecture notes to another language, or batch-translate Markdown documents.

Convert math-heavy PDFs (lecture notes, textbooks, papers, problem sets) into clean Markdown with proper LaTeX math. Use this skill whenever the user asks to convert a PDF containing mathematics to Markdown, extract formulas to LaTeX, make PDF course materials searchable/editable, or batch-process math PDFs for use in note-taking or LLM applications.