CloseCrab

NVIDIA NVL72 (GB200/GB300) GPU 集群验收测试。Profile 驱动的端到端质检：硬件自检、DCGM 诊断、NCCL 带宽、cuBLAS GEMM、单域/跨域多节点 NCCL、日志分析、报告生成、故障节点处理。含验收 baseline 阈值与判定标准。当用户说"验收测试"、"GPU QA"、"质检"、"跑一下 hw-check"、"NCCL benchmark"、"节点健康检查"、"GPU 集群测试"、"NVL72 测试"、"GB300 QA"、"GB200 QA"等关键词时触发。

Build polished, on-brand slide decks (python-pptx → Google Slides) and richly formatted Google Docs, with a reusable design system, matplotlib charts, AI-generated imagery (Nano Banana), a render-verify loop, and link-stable Drive uploads (PATCH same fileId so the share link never changes). Use when the user wants to create or iterate a presentation / briefing deck / exec one-pager / formatted Google Doc — especially customer or executive briefings that go into Google Drive. Also covers extracting user-pasted images from the session transcript and converting decks/docs to PNG for visual self-check.

Generate music using Google Lyria 3 (DeepMind). Use for "生成音乐", "作曲", "写一首曲子", "generate music", "create a song", "compose", "piano piece", "钢琴曲", "make a beat", "生成钢琴曲", "来一首歌".

Read WeChat Official Account (微信公众号) articles. Trigger when user shares any mp.weixin.qq.com URL, or says "微信文章", "公众号文章", "读一下这个微信链接", "WeChat article". Bypasses WeChat's anti-scraping CAPTCHA by emulating the WeChat in-app browser. MUST be used instead of WebFetch/Jina for any mp.weixin.qq.com link — those tools WILL fail with "环境异常" CAPTCHA.

Compute Hong Kong salaries tax partial exemption for share awards (GSU / RSU / stock options) under DIPN No. 38 paragraphs 43-46 and section 8(1A)(a) of the Inland Revenue Ordinance. Full lifecycle: (1) auto-extract vest data from Morgan Stanley / IRD / IR56B via Chrome MCP, (2) compute DIPN 38 time apportionment, (3) cross-check IR56B vs IRD assessment to detect unclaimed exemptions, (4) generate APPENDIX-EQUITY PDF for eTAX submission, (5) generate HTML audit report for personal review, (6) guide eTAX BIR60 filing or 70A retrospective claim. Triggers: HK salaries tax, BIR60, IR56B, share award apportionment, DIPN 38, 'partial exemption', cross-border GSU/RSU vest, '报税分析', '股票减税', 'GSU 豁免', '从内地转来香港', '70A', 'time apportionment', '股票报税'.

Guide for selecting and configuring distributed training strategies in NeMo AutoModel, including FSDP2, Megatron FSDP, DDP, and parallelism settings.

Guide for onboarding new model architectures into NeMo AutoModel, including architecture discovery, implementation patterns, registration, and validation.

Create and modify NeMo AutoModel training and evaluation recipes, including YAML structure, builders, and execution flow.

Validate and use selective and full activation recompute in Megatron Bridge to reduce GPU memory usage at the cost of extra compute.

Validate and use CUDA graph capture in Megatron Bridge, including local full-iteration graphs and Transformer Engine scoped graphs for attention, MLP, and MoE modules.

Validate and use MoE expert-parallel communication overlap in Megatron-Bridge, including overlap_moe_expert_parallel_comm, delay_wgrad_compute, and flex dispatcher backends such as DeepEP and HybridEP.

Techniques for reducing peak GPU memory in Megatron Bridge — expandable segments, parallelism resizing, activation recompute, CPU offloading constraints, and common OOM fixes.

MoE expert-parallel communication overlap in Megatron Bridge. Covers dispatch/combine overlap, flex dispatcher backends, and expert wgrad scheduling.

Validate and use packed sequences and long-context training in Megatron-Bridge, distinguishing offline packed SFT for LLMs from in-batch packing for VLMs, and applying the right CP constraints.

Recommend and customize Megatron Bridge recipes for a user's model, GPU count, and training goal. Indexes library recipes (pretrain/SFT/PEFT) and performance recipes.

Converts cuTile GPU kernels (@ct.kernel) to Triton (@triton.jit). Handles standard in-repo conversion, debugging (cudaErrorIllegalAddress, shape mismatch, numerical mismatch), and mapping cuTile idioms (ct.load/ct.store, ct.Constant, ct.launch) to Triton equivalents. Covers dual-kernel layout flags (e.g. transpose=True/False + autotune grid via META) per translations/advanced-patterns.md. Use when converting, porting, or translating cuTile kernels to Triton, or debugging existing Triton translations.

Integrate TileGym kernels into Hugging Face `transformers` models by replacing the library's submodule(s) and certain class(es)' implementations, and patching certain class(es)' init/forward/load weight methods prior to instantiating models. Used when the user requires integrating TileGym kernels into `transformers` models.

Generate speech audio from text using Gemini 3.1 Flash TTS (with emotion control) or Edge TTS fallback. Use when the user says "/tts", "语音合成", "文字转语音", "念出来", "读出来", "生成语音", "帮我念", "说出来", or when asked to convert text to speech audio.

香港智能通勤助手。Google Maps 路线推荐 + KMB/Citybus 实时到站融合。当用户说"几点有车"、"巴士到了吗"、"怎么去公司"、"坐什么车"、"查公交"、"bus ETA"、"下一班车"、"通勤"、"commute"等关键词时触发。

以 Chris 身份通过飞书 API 发消息到 bot DM 窗口。支持文字和语音。Siri 快捷指令的服务端。

CC Pages 增量备份到 GitHub private repo

Real-time SVG whiteboard for teaching. Draw shapes, arrows, text step-by-step while explaining concepts. Viewers watch on any browser. Use when giving visual explanations, drawing architecture diagrams, or teaching algorithms.

聊天平台消息格式化规则。当 Claude Code 通过聊天平台（Discord、飞书等）运行时自动适配消息风格，避免渲染问题。当用户说"聊天风格"、"chat风格"、"消息格式"时触发。

查天气预报。香港用香港天文台官方 API（最准、秒回），全球用 Open-Meteo（免 key）。当用户说"天气"、"天气预报"、"今天天气"、"明天天气"、"几天天气"、"会不会下雨"、"weather"、"forecast"、"香港天气"、"某城市天气"等关键词时触发。

Generate a comprehensive multimodal HTML explainer document with embedded Gemini TTS voice narration segments, structured sections, callouts, diagrams, and tables. Optimized for explaining complex technical topics (PoC briefs, new domain concepts, architecture deep-dives) to first-time learners who need both written reference AND spoken walkthrough. Use when the user says "做个讲解文档"、"多模态讲解"、"配语音讲解的文档"、"科普文档加语音"、"explain this with voice"、"technical walkthrough"、"语音讲解 HTML"、"多模态科普", OR when a topic is complex enough that an audio walkthrough materially helps comprehension (50+ concepts, multi-stakeholder, first-time exposure).

Fill a .docx template (forms, portfolios, application sheets, report cards) with data extracted from a source document (PDF, another docx, chat input) while preserving the template's formatting. Use when the user provides a .docx template plus a data source and says "帮我填这个表"、"按这个 PDF 填模板"、"把这份资料填进 docx"、"fill this form"、"fill this template from a PDF"、"portfolio 填表"、"申请表填一下". Handles paragraphs, table cells, AND textbox/cover-page content (which python-docx alone cannot reach).

飞书企业邮箱管理与收发。当用户说"发邮件"、"收邮件"、"查邮件"、"回复邮件"、"send email"、"check email"、"加个邮箱"、"配置邮箱"、"create mailbox"、"setup email"等关键词时触发。

全面 memory 重构战役 — 不只 audit，是把"2026-05-22 战役"完整流程做一遍（健康审计 → 集群合并 → 孤儿处理 → 死链修复 → 小爱回归测试 → 备份）。当用户说「memory lint」「/memory lint」「记忆大扫除」「memory 大整理」「memory 重构」时触发。跟 memory-gc skill 互补 — memory-gc 只 audit，memory-lint 是大手术。

记忆系统自动 GC — audit 当前 MEMORY.md / shared/ / disk memory 文件健康度。**Pressure-driven 不是 time-driven** — 没积累就不该清, cold ≠ garbage, 没新东西不要找事清。当用户说「memory audit」「记忆 GC」「记忆清理」「sleep gc」「/memory gc」「/memory audit」「check memory health」时触发。

Bot team three-way mutual worker-optimization loop. Two evaluator bots design and dispatch test cases to a target bot via Firestore inbox, monitor real-time logs, compute metrics (fail_rate, empty_response_count, p50/p95 duration, avg_step_count), diagnose root causes, propose source-code fixes, then trigger cross-bot SIGHUP restart and re-test — all autonomously without asking Chris. Use when user says "进化"、"evolve"、"evolution round"、"进化一轮"、"start a round"、"互相 restart"、"三方互评"、"今晚优化 <bot>"、"组队优化" or asks the bot team to autonomously improve a specific worker (ClaudeCodeWorker / KiloWorker / OpenClawWorker / GeminiACPWorker).

CloseCrab bot 部署/重启后健康检查。当用户说"体检"、"smoke test"、"health check"、"检查 bot"、"bot 还好吗"、"刚部署完确认下"或部署/重启后想验证状态时触发。覆盖进程、Firestore、worker secrets、日志活性、近期错误，并扫描 `~/.closecrab/smoke-tests.d/` 下用户自定义 drop-in 检查。

Generate or edit images using Gemini 3 Pro Image (Nano Banana) on Vertex AI. Use for text-to-image ("生成图片", "画一张", "generate image", "帮我画", "create an image", "图片生成") and image-to-image editing ("修图", "P图", "改图", "edit image", "在这张图上加 ...", "基于这张图 ...", "保持风格生成 ...").

First-pass codebase reconnaissance using Google Code Wiki (codewiki.google). Use when the user is opening a new repository for the first time, asks for a quick architectural overview of an unfamiliar codebase, or needs to build a mental model before diving into code. Pulls the auto-generated TOC, architecture diagram, and section summaries from Code Wiki to bootstrap understanding. Strictly a recon tool, not a Q&A oracle — Code Wiki cannot answer deep runtime/optimization questions, only describe architecture.

Generate teaching video explanations for elementary math problems. Takes a math problem (with answer and solution steps) and produces a 2-3 minute MP4 video with dark-themed slides and TTS narration, suitable for 5th graders. Use when asked to create math teaching videos, explain math problems to kids, or generate video tutorials for math exercises. Supports batch processing for multiple questions.

大白话论文解读 - 将学术论文翻译成通俗易懂的中文解读文档，配有图解、示例代码和知识点补充

Generate a handoff prompt that lets a fresh Claude Code session resume a complex task from a crashed/restarted CloseCrab bot session. Reads the old session jsonl plus the bot's recently-written CC Pages docs, extracts task identity / latest state / file locations / pitfalls / team coordination signals, and synthesizes a concise markdown prompt to paste into the new session. Use when a CloseCrab bot's session crashed (API error, context overflow, OOM-killed), or when the user says "wrap up this context for a new session", "give me a handoff prompt for {bot}", "{bot} 的 session 挂了写个交接", or "总结一下让新 session 接着干".

Use when the user wants to create any technical diagram - architecture, data flow, flowchart, sequence, agent/memory, or concept map - and export as SVG+PNG. Trigger on: "画图" "帮我画" "生成图" "做个图" "架构图" "流程图" "可视化一下" "出图" "generate diagram" "draw diagram" "visualize" or any system/flow description the user wants illustrated.

This skill should be used when users need to install the Lustre client and mount a Google Cloud Managed Lustre filesystem on GCE VMs (especially TPU VMs). It covers client installation via Artifact Registry, kernel module loading, mounting, and network troubleshooting.

Generate videos using Google Veo 3.1 on Vertex AI. Use when the user says "生成视频", "做个视频", "generate video", "帮我生成一段视频", "create a video", "视频生成", "text to video", "文生视频", or when you need to create video content.

Create stunning, animation-rich HTML presentations from scratch or by converting PowerPoint files. Use when the user wants to build a presentation, convert a PPT/PPTX to web, or create slides for a talk/pitch. Helps non-designers discover their aesthetic through visual exploration rather than abstract choices.