SuperAPI AI Directory

university-textbook

General & other

View repository

Updated 10-16 09:30

About this skill

University Textbook · About 20 pages, quick reading

By default, deliver a concise textbook of about 20 pages: converge the topic scope, prioritize clearly explaining key principles and one complete example. Retain the derivations necessary for understanding; do not compress the page count by shrinking the font size. If the user specifies a length, their requirement takes precedence.

This skill distills the world-class full-process specification for writing deep learning and embodied intelligence robotics textbook tutorials, aiming to solve bottlenecks in writing long-form academic lecture notes such as "compressed length, skipped derivation steps, abbreviated code, and layout overflow," and to achieve high-quality long-form tutorial output that combines the inspirational power of Feynman-style plain language, a high-school-level math and science cognitive scaffold, microscopic purely manual calculations, runnable end-to-end code, and rigorous mathematical appendices.


The Four Core Pillars and Standard Specifications

1. Tightly interlocking causal chain across chapters (The Narrative Thread)

  • Section 1: cognitive origins, physical common sense, and underlying contradictions (point-mass model / pixel entropy trap / memory wall / Moravec's paradox);
  • Section 2: analysis of classic baseline architectures and the crisis of geometric/dynamical collapse (Representation Collapse, autoregressive cumulative drift);
  • Section 3: core mathematical breakthroughs of the new paradigm and dynamical intuition (asymmetric dynamics, Stop-Gradient, EMA momentum flywheel, continuous flow matching);
  • Section 4:: microscopic single-step purely manual calculation (extremely critical): set low-dimensional numerical values (e.g., $D=2$), and calculate every real number in the forward prediction, loss, backpropagation subgradient, parameter update, and EMA smoothing without skipping a single step;
  • Section 5: provideacademic-level LaTeX pseudocode (must strictly follow the algorithm + algorithmic specification), and writea directly runnable end-to-end PyTorch Demo network, accompanied by torch.allclose zero-error verification against the official industrial implementation;
  • Section 6: panoramic generational comparison table + industrial-grade deployment pitfall-avoidance guide + top-conference paper study navigation map;
  • Appendix (Ladder 0 -> 1 -> 2): completely strip away the hardest mathematical foundations (Lie group manifolds, continuous-limit ODEs, Lyapunov stability, VICReg covariance full-rank theorem) and explain them thoroughly step by step.

2. High-school math and science cognitive scaffolding (Cognitive Scaffolding)

  • Ladder 0 (daily life / high school common sense): use the most plain everyday phenomena to build physical intuition (e.g., a painter capturing the spirit of a horse, muscle memory when swinging a tennis racket, inertial pull in pairs figure skating);
  • Ladder 1 (low-dimensional algebraic hand calculation): use two-dimensional coordinates $[x, y]$ and single-variable equations with actual concrete numerical values substituted in to calculate each step;
  • Ladder 2 (high-dimensional rigorous functionals and original top-conference definitions): provide the original formulas and rigorous proofs from top-conference papers.

3. Iron rules for TikZ vector figures to prevent overflow and for precise layout

  • The full text contains at least 4 high-precision TikZ figures;
  • single-node encapsulated layout: text inside a box must be written within the same Node's {} It is strictly prohibited to use loose text that causes text to overflow the border;
  • explicit text width constraint: for any node with minimum width=W set, it must explicitly declare text width=W-0.5cm, align=center, and configure inner sep=4pt~6pt;
  • connector text collision prevention: if there is text annotation between two modules, the horizontal clearance must be reserved as $\ge 2.0\text{cm}$, and the text should be laid out with multiple lines and wrapped;
  • the overall canvas width should be controlled within 12cm (aroundscale=0.9 to eliminate Overfull \hbox。

4. Academic-level LaTeX pseudocode standard specification (Strict Aunt Er / official standard)

In LaTeX documents, pseudocode layout follows the following standard specifications:

\usepackage{algorithm}
\usepackage{algorithmic}

\begin{algorithm}[htbp]
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\caption{算法标题}
\label{alg:my_algorithm}
\begin{algorithmic}[1]
\STATE \textbf{Input:} 输入变量与超参数
\STATE \textbf{Output:} 输出结果
\STATE 初始化参数
\REPEAT
    \STATE 核心计算与前向传播
    \FOR{$i = 1$ \TO $M$}
        \STATE 单步更新与条件分支
    \ENDFOR
    \STATE 反向传播与梯度更新
    \STATE 动量平滑演化
\UNTIL 收敛
\STATE \textbf{Ensure:} 最终优化模型
\end{algorithmic}
\end{algorithm}

Multiple-Stage Chunked Pipeline execution strategy (Multi-Stage Chunked Pipeline)

When writing a rapid-reading textbook of about 20 pages, divide it into chunks by content and control the length of each part, adopting multi-stage pipeline scheduling:

  1. Stage 0: Global state and symbol registry(lock down the unified symbols and reference interfaces for the whole book);
  2. Stage 1: Chunked in-depth writing(Chunk A: Sections 1-2 $\to$ Chunk B: Sections 3-4 $\to$ Chunk C: Section 5 code $\to$ Chunk D: Section 6 comparison $\to$ Chunk E: Appendix derivation);
  3. Stage 2: Stitching and cross-reference link repair;
  4. Stage 3: Automated double compilation and TikZ layout self-healing。

Supporting resources

These files are located in the same skill directory as SKILL.md and must be retained together during installation.