Convert LaTeX manuscripts (especially ctex / ctexart Chinese academic papers) to editable Word (.docx) while preserving cross-references for figures, tables, equations, and bibliography numbering, and while matching a target journal template's styling. Use this skill whenever the user wants to export a .tex file to .docx, convert a LaTeX paper to Word, run pandoc tex→docx, produce a Word version of a manuscript for a Chinese journal (气象与环境学报 etc.), or fix a broken LaTeX→Word export (missing figure list, broken cross-refs, wrong numbering, ctex layout commands leaking through, equation refs not clickable, template styling not applied). Trigger even when the user just says "turn this paper into Word" or "I need a .docx of this manuscript" without mentioning pandoc.
Construct, verify, and communicate mathematical proofs for ML theory — convergence rates, generalization bounds, architecture properties, optimization analysis, generative model theory, and scaling laws. Use when the user asks to prove a theorem about an ML algorithm, verify a convergence bound, analyze Lipschitz properties of a network, derive generalization guarantees, prove properties of diffusion models or flow matching, analyze NTK behavior, construct universal approximation arguments, or build any rigorous mathematical argument in the context of machine learning. Also trigger on: convergence proof, generalization bound, PAC-Bayes, Lipschitz analysis, gradient flow proof, lower bound, concentration inequality, optimization theory, induction on iterations, equivariance proof, score matching, diffusion SDE, flow matching, NTK, neural tangent kernel, Rademacher complexity, Wasserstein distance, optimal transport, scaling law, universal approximation, loss landscape, PL condition.
PyTorch model design, computation graph analysis, profiling, and optimization — from architecture implementation to production-ready training and inference. Use when the user asks about implementing neural architectures in PyTorch, profiling FLOPs/memory/throughput, optimizing training speed or memory usage, debugging gradient flow or numerical stability, choosing between dynamic and static graphs (eager vs torch.compile), setting up distributed training (FSDP2, tensor parallel, pipeline parallel, 2D/3D parallelism), multimodal or temporal model architecture, or optimizing inference (quantization, KV cache, speculative decoding, continuous batching). Also trigger on: computation graph, autograd, torch.compile, FSDP2, DTensor, tensor parallelism, pipeline parallelism, mixed precision, AMP, bf16, fp8, gradient checkpointing, FlashAttention, model profiling, OOM debugging, CUDA memory, inference optimization, model quantization, torchao, multimodal, cross-attention, vision encoder, temporal model.
Hypothesis-driven research workflow for turning research ideas into auditable study plans, experiment designs, execution tracking, and evaluation reports in markdown. Use this skill whenever the user is managing an AI or ML study across framing, experiment design, execution, evaluation, or research iteration; checking gate readiness; tracking baselines, ablations, or evidence; or maintaining a research roadmap and audit trail. Prefer this skill for multi-step research management tasks that need reproducibility, explicit state, and validation across sessions. Works with any AI agent tool that can read files, run commands, and edit markdown.