dotfiles

Review committed and working-tree changes since a fixed point against repository standards, the originating spec, and correctness or risk. Use when the user asks to review a branch, pull request, work in progress, current changes, or changes since a commit, branch, tag, or merge base.

Relentlessly clarify a fuzzy engineering or product plan one decision at a time while maintaining durable domain language and architectural decisions. Use when the user wants to grill, stress-test, think through, or sharpen a plan, design, feature, trade-off, or architecture choice before writing a spec or code.

Implement the user's current engineering task with minimal, verified changes, test-driven discipline, and a final code review. Use when the user asks Codex to build, fix, implement, modify code, execute a concrete task, or complete work from specs/slug/issues.md.

Break a product spec, technical spec, design note, or current plan into one or more dependency-aware, agent-executable issue briefs under specs/slug/issues.md. Use when the user asks to split work into issues, implementation tasks, tickets, milestones, TODOs, or independently verifiable slices.

Convert current context, repo findings, or an existing thinking note into durable product and technical specification markdown under specs/slug/. Use when the user asks to write, generate, update, revise, or turn a plan or discussion into a spec, PRD, product requirements document, or technical proposal.

发布前对应用程序进行全面审计，检查安全性、数据泄露、可靠性、并发状态完整性、可访问性和 UI 一致性。适用于上线、发布或交付前的安全检查。

Convert a finished spec or design document into a right-sized execution plan and /goal starter for Codex. Use when the user has a completed spec/design and wants phases, a phased plan, an execution checklist or JSON ledger, a progress file, hard acceptance criteria, batch-processing lists, or a /goal prompt. Triggers include "基于这个 spec 写 /goal plan", "把设计文档拆成可执行 phase", "生成 phased plan 和 progress 文件", "逆向清单做成 checklist 分批处理". Routes small tasks to a lighter ledger instead of full phases. Not for brainstorming, writing the original spec, or implementing the plan.

Detect software architecture bad smells, algorithmic complexity hotspots, and anti-patterns in a codebase. Produces a detailed markdown report identifying violations of architectural principles, design patterns, code quality, and performance complexity. Triggers on: smell, code smell, architecture smell, find anti-patterns, detect bad smells, complexity analysis, 代码坏味道, 架构坏味道, 反模式, 找出坏味道, 复杂度分析.

Find deepening opportunities in a codebase, informed by the domain language in CONTEXT.md and the decisions in docs/adr/. Use when the user wants to improve architecture, find refactoring opportunities, consolidate tightly-coupled modules, or make a codebase more testable and AI-navigable.

Reproduce and diagnose bugs, test failures, flaky behavior, regressions, crashes, wrong outputs, or performance problems before fixing them. Use when the user says something is broken, failing, throwing, slow, flaky, regressed, or asks to debug, investigate, reproduce, find root cause, or verify a suspected fix.

Analyze and improve codebase architecture or maintainability. Use when the user asks to improve, refactor, simplify, deepen, decouple, reduce technical debt, make code easier to test, or review architecture and module boundaries.

Teach the user a topic or skill through mission-grounded lessons, trusted resources, tight feedback loops, and durable learning records. Use when the user asks to learn, study, understand, practice, or be taught something.

Give a higher-level map of an unfamiliar code area, including relevant modules, callers, boundaries, and domain vocabulary. Use when the user asks to zoom out, understand how code fits together, or get broader context before changing code.

Expert-level knowledge about AI/ML knowledge distillation — a training paradigm where a student model learns from a teacher model. Use this skill whenever the user asks about knowledge distillation, teacher-student learning, model compression via distillation, soft targets, KD loss, on-policy distillation, off-policy distillation, self-distillation, multi-teacher distillation, sequence-level KD, GKD (Generalized Knowledge Distillation), OPD (On-Policy Distillation), MOPD (Multi-Teacher On-Policy Distillation), OPSD (On-Policy Self-Distillation), DistilBERT, soft labels, temperature scaling, KL divergence in distillation, or any RL–distillation hybrid method. Also trigger when asked about compressing large LLMs into smaller ones, transferring capabilities from frontier models, exposure bias in autoregressive models fixed via student rollouts, or synthetic data pipelines for model training.

SIMD (Single Instruction Multiple Data) 优化技巧知识库，蒸馏自 Daniel Lemire 博客（lemire.me/blog） 的数百篇性能工程文章与研究论文。覆盖 x86 (SSE/AVX2/AVX-512) 和 ARM (NEON/SVE2) 的实战优化 模式、代码范例与性能直觉。当用户提问涉及 SIMD 向量化、性能优化、数据并行、字节解析、 JSON/文本/UTF-8/base64 加速、位图操作、整数压缩、prefix sum、字符串扫描、 类型转换加速时触发。也适用于"为什么我的循环慢"、"如何用 intrinsic 优化"、"AVX vs NEON" 等问题。只要涉及批量数据处理性能就应优先查阅本 skill。

使用 Nsight Compute 对 CUDA kernel 进行性能分析和优化的完整指南。 当用户提到 CUDA 性能分析、kernel 调优、ncu 命令、profiler 数据解读、 瓶颈诊断、SOL 指标、占用率分析、DRAM Throughput、L1/L2 Cache Throughput、 Compute Throughput、Roofline 图、Warp Stall、Memory Chart、Section Sets、 Scheduler Statistics、Warp State、Branch Divergence、Register Spilling 等话题时， 主动使用此 skill。也适用于用户粘贴 ncu 输出数据需要解读的场景。 资料来源：NVIDIA 官方文档 2026.1、Kernel Profiling Guide、NERSC Roofline 文档。

GPU 术语速查手册，基于 Modal GPU Glossary（CC BY 4.0）。当用户提问涉及 GPU 硬件结构、CUDA 编程模型、GPU 性能分析、内核优化、内存带宽、屋顶线模型、 SM 占用率、Warp 分化、算术强度等概念时触发。也适用于"为什么我的 kernel 慢"、 "compute-bound 和 memory-bound 怎么区分"等实际调优问题。

语音推理优化参考技巧库。当用户提问与语音推理、语音 AI 性能、实时语音系统延迟、ASR/TTS 优化相关时触发。包含 RTF、流式 pipeline、VAD、量化、模型选型等实战技巧。即使用户只是问"怎么让语音更快"、"TTS 延迟太高怎么办"、"ASR 流式怎么做"也应触发。