self-skills

DeepSpeed ZeRO 显存优化与训练加速

Flash Attention 版本选型与最佳配置

提交代码到远程仓库

NVIDIA Nsight Systems 系统级 GPU Profiling

Nsight Systems 自动检测、安装与升级

PyTorch Profiler 性能分析，提供算子级瓶颈定位

外网代理配置，支持受限网络环境中访问 GitHub/PyPI 等外网资源

机器资源信息采集（CPU/内存/GPU/磁盘/网卡）

LLM/MLLM 训练加速审计，扫描项目并输出评分报告

Triton 算子性能优化与热点算子融合加速

端到端 GPU 训练智能调优 — 静态分析 → 硬件采集 → Profiling → 瓶颈诊断 → 代码优化 → 验证迭代

纯静态代码分析，无需 GPU/运行环境，扫描训练项目并输出优化建议报告

Flow-Factory 训练优化审计

LlamaFactory 训练优化审计

SGLang 推理优化审计

ms-swift 训练优化审计

HuggingFace Transformers Trainer 训练优化审计

VideoX-Fun 训练优化审计

vLLM 推理优化审计

Use before research or engineering-research decisions: deciding whether to continue a direction, run an experiment, change route, interpret results, write a paper claim, or invest more work when there is risk of local optimization, self-drawn targets, weak evaluation, LLM self-evaluation loops, system-specific bug fixes disguised as innovation, or poor ROI. Produces a concise decision gate covering generality, innovation boundary, evaluation credibility, ROI and stop conditions, paper narrative fit, next action, and what not to do.

多 Agent 协作议会 - 主 agent 派遣多个 sub-agent 对同一问题独立给出完整答案，主 agent 担任裁判+合成器，输出对比分析后的最终答案。灵感来自 OpenRouter Fusion。适用于需要多视角、深度调研、避免单模型偏见的复杂问题。触发词：多模型协作、模型议会、多视角分析、llm council、fusion、议会讨论、几个 AI 一起讨论。

生成一份自包含的决策上下文闭包，用于向不能访问本地文件的外部强 AI 求助。当用户觉得当前 agent 不够聪明、想找网页端 AI/外部专家求助、需要决策/认知交接而非代码交接、想要外援请求包/决策上下文包、或需要把项目上下文、历史认知修正、约束条件、基建条件、未解决问题整理成外部可消费格式时使用。

AI 研究顾问：自动判断用户问题是「方向选择」还是「实验设计」，然后动态调度多个子 agent 并行分析，最后整合输出。 当用户提到「纠结方向」「值不值得做」「选哪个」「实验怎么设计」「怎么验证」「baseline 怎么设」「帮我看看这个研究方向」「AI 顾问」时触发。 即使只是普通地讨论算法改进或研究困惑，也应自动调用此 skill。

顶级会议论文搜索推荐 - 搜索 CVPR/ICCV/ECCV/ICLR/AAAI/NeurIPS/ICML 等顶会论文

从论文中提取图片，优先从arXiv源码包获取真正的论文图

女娲造人：输入人名/主题/甚至只是模糊需求，自动深度调研→思维框架提炼→生成可运行的人物Skill。 两种入口：(1)明确人名→直接蒸馏 (2)模糊需求→诊断推荐→再蒸馏。 触发词：「造skill」「蒸馏XX」「女娲」「造人」「XX的思维方式」「做个XX视角」「更新XX的skill」。 模糊需求也触发：「我想提升决策质量」「有没有一种思维方式能帮我...」「我需要一个思维顾问」。

深度分析单篇论文，生成详细笔记和评估，图文并茂 / Deep analyze a single paper, generate detailed notes with images

Forces high-agency exhaustive problem-solving with corporate PUA pressure. Triggers on user frustration, repeated failures (2+), passive behavior, or quality complaints. Common triggers across Reddit/LinuxDo/HN/X: 'try harder', 'figure it out', 'stop giving up', 'you keep failing', '加油', '别偷懒', '你再试试', '为什么还不行', '你怎么又失败了', '你怎么搞的', '又错了', '能不能靠谱点', '认真点', '不行啊', '降智了', '你又在原地打转', '你把之前的改坏了', '别让我手动处理', '换个方法', 'stop spinning', 'you broke it', 'why does this still not work', 'this is the third time', '/pua', 'PUA模式'. Applies to ALL task types: code, config, debug, deploy, research.

论文阅读工作流启动 - 生成今日论文推荐笔记 / Paper reading workflow starter - Generate daily paper recommendations

所有联网操作必须通过此 skill 处理，包括：搜索、网页抓取、登录后操作、网络交互等。 触发场景：用户要求搜索信息、查看网页内容、访问需要登录的网站、操作网页界面、抓取社交媒体内容（小红书、微博、推特等）、读取动态渲染页面、以及任何需要真实浏览器环境的网络任务。

Governs Copilot Agent interaction continuity in all conversations. Use this skill whenever a task or subtask is about to be marked as complete, a question needs clarification, or a workflow phase ends. This skill MUST be active in every coding session — it prevents the agent from abruptly ending the conversation and enforces active follow-up after every piece of completed work.

Business-oriented SOP for optimizing algorithm/AI engineering resumes targeting top-tier tech company campus recruiting or internship positions. Use whenever a user wants to write, rewrite, review, or polish a resume entry — whether it's a project, a paper, or a competition. Also trigger when the user provides raw project notes, interview reflections, paper drafts, competition summaries, or draft bullet points and wants to turn them into compelling resume content. Trigger proactively when the user shares a .tex resume file and asks for feedback, edits, or review.

Specialized skill for compiling, validating, and iteratively refining LaTeX resumes. Use this skill whenever a .tex file is modified (content or layout), a LaTeX resume needs to be compiled, or PDF output needs to be verified. Automatically triggers after ANY edit to a .tex file in a resume workspace — enforce this even when the user only asks for a "small change".

Prepare and verify SSH access to remote Linux servers before editing remote code or running experiments. Use when Codex needs to read local SSH config, inspect existing keys, reuse Tabby or local SSH settings, install a public key from an already-open server terminal, debug password/key authentication failures, or confirm non-interactive SSH access to a named host such as `3090` or `host132`.