agent-skills

巡检本期 SSH 可连的 5 台远程服务器（HPCC / BCC / TASL-LabServer / TASL-7 / Brev）的 GPU 情况——每卡显存与利用率、Slurm 集群的账号配额/排队作业/全节点空闲 GPU，并算出"我实际还能新申请多少张"。当用户说"检查 GPU""看下 GPU 显存""GPU 情况""哪台服务器有空卡""查一下算力""where can I run""check GPU status""GPU 够不够""有没有空闲显卡"等，或在准备起训练/挑服务器时触发。

统一绘图 skill——文生图 / 图生图 / 多图合成 / 编辑 / 论文与科普配图。两个后端由用户指定：nano(默认，Google Gemini "Nano Banana"，快 3-10s、便宜、原生中英文文字渲染、可 1K/2K/4K，含 lite/nb2/pro 三档) 或 gpt(速创API GPT-Image-2，真实感/精细文字更强，异步稍慢 60-90s)。四个专用风格预设也由用户按名指定，两后端通用：UniVLA(pastel macaron 论文方法图)、Physical Intelligence(π/pi-zero retro-computing monospace 图)、Fast-WAM(CVPR/ICLR 矢量图+attention-mask 矩阵)、蓝皮书 BlueBook(蓝白科技科普/咨询长图)。Use this skill for ANY image generation or editing request. 触发词：「画一张」「生成一张图」「出张图」「文生图」「图生图」「改这张图」「编辑图片」「把这张图改成…」「合成/融合这几张图」「做张海报」「配图」「方法图」「架构图」「用 nano/gemini 画」「用 GPT 画」「generate/edit/compose an image」「make a poster」「render a figure」，以及风格名「UniVLA 风格」「Physical Intelligence / PI / π 风格」「Fast-WAM 风格」「蓝皮书 / BlueBook 风格」。默认走 nano；用户点名「用 GPT」时走 gpt。

Write structured experiment reports from experiment-run output. Consumes results/<run_id>/ + EXPERIMENT_TRACKER.md. Reports are scannable: TL;DR + summary table first. Trigger when user says "写个报告", "总结实验结果", "实验报告", "experiment report", or after experiment-run completes.

Adversarial paper review + improvement loop. Sends paper PDF/manuscript to Codex GPT-5.5 as a senior ML reviewer, scores 7 dimensions, then iteratively fixes weaknesses until score ≥ 8. Use after paper-write, or when user says "improve paper", "审稿", "review my paper", "polish paper", "paper review", "提升论文质量", "review loop".

Write a structured rebuttal using Concede–Clarify–Commit structure. Parses reviewer comments, identifies overlapping concerns across reviewers, builds consolidated Table R1, produces venue-formatted rebuttal. Use when user says "写 rebuttal", "rebuttal", "回复审稿意见", "response to reviewers", or receives paper reviews and needs to respond.

Execute experiments from EXPERIMENT_TRACKER.md produced by experiment-plan. For each TODO run: launch training/eval, monitor via gpu-train-monitor, auto-diagnose failures via train-debug, collect results, update tracker. Bridges the gap between planning (experiment-plan) and reporting (experiment_report). Use when user says "跑实验", "run experiments", "执行实验", "按照 plan 跑", or has an EXPERIMENT_TRACKER.md ready to execute.

Automated full-spectrum research survey. Uses litian-academic-search (7-source) for academic papers + WebSearch site: filters for social media. Ingests all found papers via paper-read --light. Produces structured FINAL_REPORT.md per topic + concept glossary with pseudocode. Use when user wants "全面调研", "综合调研", "survey this topic", "调研一下", "帮我全面了解一下这个方向".

Turn a refined research proposal into a claim-driven experiment roadmap with progressive 4-stage execution order. Reads from research-refine output (FINAL_PROPOSAL.md). Produces EXPERIMENT_PLAN.md + EXPERIMENT_TRACKER.md. Use after research-refine, or when user says "实验计划", "experiment plan", "设计实验", "ablation matrix", "evaluation protocol", "run order".

Full paper writing orchestration — from blank LaTeX project to camera-ready PDF. Delegates to paper-intro-writing (Introduction), paper-related (Related Work), and reads research-refine/experiment-plan outputs. Use when user says "写论文", "write paper", "开始写 paper", "paper writing", "draft paper", or wants to produce a submission-ready paper from experiment results.

Turn a vague research direction into a problem-anchored, concrete method proposal with claim-driven validation. Integrates litian-academic-search (literature grounding), paper-related (citation context), and gap-to-method (design space positioning). Use when user says "refine my approach", "帮我细化方案", "打磨idea", "这个想法可行吗", "refine research plan", or has a fuzzy direction that needs to become a specific, testable research proposal ready for experiment-plan.

Given a paper (arXiv ID / URL / title), find its related works through three complementary lenses: (1) citation graph — predecessors (references) and successors (citations) via Semantic Scholar + OpenAlex, (2) semantic similarity via litian-academic-search, (3) shared references/citations analysis. Outputs a structured related-work map suitable for paper writing or literature review. Use whenever the user wants to find related papers, expand from a seed paper, do citation snowballing, "找相关论文", "这个方向还有哪些工作", "citation graph", "related works", or needs to map the literature around a specific paper.

Deep-read a single academic paper (arXiv ID / URL / title) and produce structured engineering reports, then DEFAULT auto-ingest into the OmniBox local knowledge base. Combines code-level deep read (arxiv-deepdive pattern: clone repo, read PDF, 01_highlevel + 02_technical with file:line) with VLM figure extraction (read-paper pattern). Use whenever the user wants to read a paper, take notes, "读论文", "paper notes", "深读这篇", "read this paper", or needs a single paper processed. Also called by orchestration skills (comprehensive-survey, topic_survey, survey-to-deck) for each paper found. The default behavior is ALWAYS ingest into OmniBox — this is the global contract: every paper found in the research pipeline gets ingested.

Survey a research field interactively, exploring papers by topic and writing a structured literature review.

Restore and KEEP UCR VPN connectivity. Bundles four fixes. (1) ROUTING fix — static routes for non-VPN hosts via the original gateway (HPCC public-IP case). (2) IPTABLES fix (Linux) — Cisco's `ciscovpn` chain ends with `DROP 0.0.0.0/0`, dropping `tailscale0` traffic; add `RETURN` rules. (3) macOS PERSISTENT auto-reconnect manager — the UCR server drops idle sessions ("Idle Timeout", b0); a keepalive (live SSH stream over the tunnel) + foreground-openconnect-held-by-expect + auto-reconnect watchdog keeps it up, with Keychain auto-login + Duo push + sudoers NOPASSWD so no typing is needed. (4) Tailscale/VPN route-hijack — after a VPN drop, Tailscale's overlapping 169.235 subnet routes steal the BCC route (HPCC still works → easy to misdiagnose). Use PROACTIVELY when user says "连了VPN就连不上", "VPN split tunnel", "tailscale 不通", "VPN 老掉线/只能连一会儿就断", "idle timeout", "VPN 自动重连/持久化/自动登录", "ssh tasl 超时但 tailscale ping 通", "ciscovpn chain dropping", "claude monitor 连不上 ssh", or has both VPN-protected and non-VPN/Tailscale

多维文献矩阵构建 → Gap 发现 → 方法提案的完整研究工作流。 给定一个研究问题和约束条件，拆分出设计维度，通过 vec-db 语义搜索 + 联网搜索定向查找论文， 构建多维交叉文献矩阵，识别空白 Gap，并基于矩阵中已有工作的实验证据提出最优方法提案。 当用户说"找 gap"、"文献矩阵"、"有没有空白方向"、"提出方法"、"gap analysis"、 "design space analysis"、"帮我分析设计空间"、"有什么新方向可以做"、 或在已有调研基础上问"下一步应该做什么方法"时触发。 也适用于用户给出约束条件后要求系统地寻找可行研究方向的场景。

Shared reference for all paper search skills. Points to litian-academic-search as the canonical multi-source search entry point (7 sources: OmniBox, arXiv, Semantic Scholar, OpenAlex, DeepXiv, WebSearch, WebFetch). This skill is NOT user-invokable — it is a reference loaded by other skills (comprehensive-survey, gap-to-method, paper_related_works, idea_refinery, topic_survey, survey-to-deck, paper-talk-deck).

把一组论文(arXiv 链接 / ID / 标题)做成一份"大 deck 串讲"形式的 Slidev HTML 演示——多篇 paper 合成一份连续幻灯片, 带封面、总览、逐篇 section、跨篇对比收尾。每篇 paper 一个 section:动机 → method 框图(复用论文原图 + 可选 Mermaid 重绘) → 核心代码片段(逐行 focus / magic-move 讲解,带 file:line 引用)→ 结果。引擎用 Slidev(代码高亮/数学/图表一等公民), 以 HTML 现场演示为主,PDF/PPTX 为备份,PPTX 可再转成 Google Slides。 复用本机已有 skill 取素材:omnibox-search(库里已有的工程报告)、arxiv-deepdive(库里没有的现产代码级报告)、 read-paper(VLM 抽 method 原图)、drawio/Excalidraw(精修框图)。 Use PROACTIVELY whenever the user wants to TURN PAPERS INTO A TALK / PRESENTATION — triggers: "把这几篇论文做成 slides / PPT 讲", "串讲这 10 篇 paper", "组会要讲这几篇", "做个论文分享/讲解的 deck", "带代码细节的论文 presentation", "paper walkthrough slides", "讲解论文 + method 框图 + 代码", "用 Slidev 把论文做成演示", "把 arxiv-deepdive 的报告做成幻灯片"。 这是"论文 → 可讲的演示"的总装 skill;若用户只想读论文出报告/笔记,用 arxiv-deepdive / read-paper; 若只要简单纯文字 Markdown 幻灯片用 marp-slide。

端到端"调研一个方向 → 深读论文 → 每篇产两份报告 → 入 OmniBox 向量语义索引 → 做成 Slidev 串讲 deck"的总编排(meta)skill。 它本身不重新实现任何一步,而是把四个已有 skill 串成一条流水线并用 Workflow 并行 fan-out: ① 发现(comprehensive-survey / 用户直接给清单)→ ② 去重 → ③ 每篇一个 agent 跑 arxiv-deepdive 写 01_highlevel + 02_technical(带 frontmatter)→ ④ update_index.py 增量入库 + search.py 验证 → ⑤ paper-talk-deck 出 deck。 Use PROACTIVELY whenever the user wants the WHOLE pipeline from a topic/paper-list to a knowledge-base + presentation — triggers: "调研 <方向> 然后读论文出报告入库再做 deck", "把这批 paper 跑全套:deep-read 进 omnibox + re-index + 出一个 deck", "深入读一两百篇 paper、每篇 high-level + technical 报告、入向量库、做 slides", "走一遍我那条调研流水线", "research a topic end to end: deep-read papers, two reports each, ingest to my KB, build a slide deck", "survey to deck", "from topic to slides", 或用户给一个主题/一组 arXiv 清单并期望"读 → 报告 → 入库 → deck"一条龙。 若用户只想要其中一步(只深读出报告→arxiv-deepdive;只入库/同步→omnibox-sync;只做 deck→paper-talk-deck;只检索→comprehensive-survey),用对应单 skill,本 skill 专做"串起全程"的编排。

Multi-source academic literature search across 7 data sources (OmniBox local KB, arXiv, Semantic Scholar, OpenAlex, DeepXiv, WebSearch, WebFetch). Use whenever the user asks about research topics, wants literature review, searches for papers, asks "what's the state of X", "find papers about Y", "文献检索", "学术搜索", "找论文", or needs to survey any academic field. De-duplicates and synthesizes results into a structured report. Fan-out by source for speed — all sources queried in parallel.

Local OCR via LightOnOCR-2-1B + llama.cpp on macOS (Apple Silicon). Use this skill whenever the user wants to extract text from images, convert screenshots to Markdown, OCR documents, or batch-process images to text. Triggers on: OCR, 图片转文字, 截图转Markdown, extract text from image, 识别图片文字, 批量OCR. The tool is ~6s per image on M4 Metal GPU, outputs clean Markdown. Already installed at ~/proj/ocr-models/ and ~/.local/bin/ocr.

Use this skill whenever the user wants to create, restyle, or polish a research or academic Slidev deck; asks to reproduce a reference slide style in Slidev; or needs clear diagrams with reliable arrows in Slidev. Pair it with paper-talk-deck or survey-to-deck when those skills generate paper/survey content, and use this skill for the visual system and diagram construction.

本地 AI 驱动的端到端视频剪辑流水线，让 Claude Code 自己把原始素材剪成成片——引擎是 FFmpeg（不依赖 DaVinci/Blender/任何付费 MCP）。覆盖：下载素材、探测、从长连续镜头里抽帧选高光（B-roll 内容选取，替代 shot detection）、mlx_whisper 语音转写 + 卡拉OK/普通字幕、librosa 音乐节拍卡点（downbeat 硬切）、暖色电影感调色、交叉转场/硬切/Ken-Burns 推镜、竖屏模糊填充、标题与下字幕、配乐淡入淡出、渲染后自检循环、project.md 跨会话记忆。Use whenever the user wants to EDIT or ASSEMBLE video from footage — 触发词：剪辑视频 / 剪个 vlog / 把素材剪成片 / 做个 montage / 旅游vlog / 卡点 / 节拍卡点 / 踩点 / 语音字幕 / 自动字幕 / 给视频加字幕 / 视频调色 / 自动剪辑 / 选素材 / edit my video / make a montage / travel vlog / beat sync / cut to the beat / auto captions / burn subtitles / color grade footage。也适用于用户给一个素材目录并希望产出一支剪好的片子。不要用于：生成全新 AI 视频（文生视频）、纯音频处理、或仅转录不剪辑（那用 audio-transcribe）。

Transcribe audio files to text using faster-whisper with smart parallel GPU processing. Trigger when user says '转录', '音频转文字', 'transcribe', '语音识别', '听写', '把这个音频转成文字', '帮我转录一下', 'speech to text', or provides audio/video files (mp3, m4a, wav, mp4) and wants text output. Also trigger when user mentions extracting text from B站 videos, podcast episodes, or meeting recordings.

Build and query a structured, provenance-tracked, freshness-aware knowledge base for any specialized domain, so an agent can pull real, citeable engineering knowledge into a long task instead of relying on parametric memory. Use this when work depends on hard-won domain details that the model is likely to get subtly wrong or out-of-date — APIs, library internals, hardware/architecture features, protocol specs, regulations, prior PRs/commits, paper results — and when you want every retrieved claim to carry a source, a date, and a confidence so it can be trusted and re-checked. Trigger it to scaffold a new knowledge base ("build a wiki of X", "index these sources"), to query one mid-task ("what does the corpus say about Y"), or whenever a long optimization/research loop keeps guessing at domain facts. Pairs with the engineering-loop and evidence-report skills.

Run a long-horizon, evidence-driven engineering loop for any task that is an experiment system, not a one-shot generation — repeatedly proposing, implementing, validating, and promoting candidates against a fixed baseline until a measurable target is met. Use this whenever a task will span many iterations and hours and needs durable state (best-so-far, failed branches, why), an immutable baseline, objective acceptance criteria, and protection against the agent quietly redefining its own success. Trigger for kernel/perf optimization, model/training/RL tuning, compiler passes, data-pipeline or query optimization, long refactors, or research where each step must be justified by evidence. Pairs with the domain-wiki (knowledge) and evidence-report (measurement) skills. Use it even when the user just says "optimize X", "beat the baseline", "iterate until it passes", or "keep improving this" without naming a framework.

Turn objective measurements into a ranked, evidence-backed diagnosis and a next-step plan — instead of guessing at what's slow/wrong and changing things hopefully. Use this whenever you need to understand why something underperforms (a slow kernel/query/endpoint/job, a regression, a model that's underperforming, high cost/memory) and decide what to fix first, grounded in profiler output, benchmark numbers, traces, or logs rather than intuition. The golden rule it enforces — Measure then Diagnose then Plan, never guess; never write a wall of suggestions, rank them by evidence times expected impact. Trigger for "profile this", "why is X slow", "diagnose this bottleneck", "read this profiler/benchmark report", "what should I optimize next", or when a long optimization loop needs to pick its next direction from data. Pairs with the engineering-loop and domain-wiki skills.

The orchestration entry point for running a Kernel-Design-Agents-style engineering loop on any domain — it ties together three sibling skills (engineering-loop for control, domain-wiki for knowledge, evidence-report for measurement) into one long-horizon, auditable, evidence-driven workflow. Use this when a task is an experiment system rather than a one-shot generation and you want the full framework, not just one layer — e.g. the user says "use the KDA approach", "set up the whole loop for this", "optimize X end to end with knowledge + profiling + iteration", or asks how the engineering-loop / domain-wiki / evidence-report skills fit together. Start here to decide which layers a task needs and in what order; this skill is thin and delegates the actual work to the three specialized skills.

在动手执行一个"长任务/大改动"之前，先和用户一起把任务的边界、目标、可验证的验收标准谈清楚，写成一份自洽的 SPEC.md，规范定义完整后才开始执行。这是一个"协作式 Plan Mode"：先润色用户的意图，用 AskUserQuestion 反向访谈补齐缺口，给出可运行的 check（测试/构建/截图对比），再放行实现。无论用户是否说"先规划"，只要任务是多步骤、跨多个文件、需要数小时自主推进、涉及迁移/重构/从零搭功能/调研，或者用户给了一个含糊的大目标（"帮我做一个 X"、"重构 Y"、"把这个项目搭起来"、"迁移到 Z"），都应主动使用本 skill 先立规范、再执行，避免 Claude 埋头解决一个错的问题。小到一句话能描述清楚的改动（改 typo、加日志、重命名）不要用本 skill，直接做。

增量拉取一个 OmniBox（小黑 / omnibox.pro/s/<token>）公开分享里"自上次以来新增或变更"的保存内容（链接/文档/文件），做最小 token 的 DIFF： 先只拉元数据比对 id + updated_at，只对 delta 拉正文，并维护一份持久化快照（全局 ~/.claude/omnibox-sync/<token>.json）。 Use PROACTIVELY whenever the user wants to check / pull / sync NEW or updated items from an OmniBox share, or says "OmniBox 又更新了", "小黑里加了新内容", "拉一下新增的保存", "看看 OmniBox 有什么新东西", "增量同步 OmniBox", "omnibox 的 diff", "omnibox.pro/s/ 新增", "sync omnibox", "pull new omnibox links", "what's new in my omnibox share"。 本 skill 单一职责：只产出"增量"（new/changed 的结构化清单 + 正文），不做下游分类/arXiv 解析/报告——那些另起（如 arxiv-deepdive）。

把 OmniBox（小黑 / omnibox.pro/s/<token>）分享里的视频资源（B站 b23.tv + 小红书 type=video）端到端"预阅读"： 下载音频 → 本地 faster-whisper(GPU) 转写 → 分类并抽取视频里口播/讨论的论文 → 与已有库去重 → 对新 arXiv 论文跑 arxiv-deepdive 出工程报告 → 归档+INDEX。 解决"视频内容无法被文本收藏覆盖"的问题（很多论文/方法只在讲座、解读视频里出现）。 Use PROACTIVELY whenever the user wants to process / transcribe / mine the VIDEO items in an OmniBox share, or says "把 OmniBox 里的视频转文字", "处理小红书/B站视频", "视频转录后抽论文", "视频预阅读", "transcribe omnibox videos", "mine papers from the saved videos", "视频里提到的论文也整理一下"。 依赖三个 skill：omnibox-sync(取视频清单)、audio-transcribe(GPU转写)、arxiv-deepdive(出报告)；本 skill 负责把它们串起来并处理下载这一最难环节。

Given a paper, find its related works (cited references and successor/citing papers) and produce a structured map.

Bootstrap a fresh Ubuntu machine to Gonglitian's personal dev setup — zsh + oh-my-zsh + fzf, Ghostty terminal (TokyoNight), Edge as default browser, VSCode, Miniconda, Claude Code with the `cc` alias + claude-hud plugin, GitHub CLI auth (HTTPS), Snipaste, WeChat, fcitx Chinese input, and Tailscale. Trigger this skill whenever the user says "setup this machine", "new ubuntu", "新电脑配置", "配一下开发环境", "装一下常用软件", "装 claude code / ghostty / edge / zsh / miniconda", "fresh ubuntu install", "new computer dev env", or asks to install any of the tools above on a fresh box. Also trigger on requests to fix Chinese input in Ghostty, to set up `gh auth login`, or to add the `cc=claude --dangerously-skip-permissions` alias.

Diagnose and fix Chinese / Japanese / Korean (CJK) input method (IME) issues in Ghostty terminal on Linux — specifically the snap-distributed Ghostty + fcitx5 + GTK4 + X11 / Wayland combination. Trigger this skill whenever the user reports: 'ghostty 输入不了中文 / 中文打不了 / 输入法不工作 / fcitx 在 ghostty 里失灵', 'ghostty 装好了但输入法没反应', 'snap 应用 + 输入法 / IME 问题', 'GTK4 + XIM 不工作', or wants to mirror a working ghostty + fcitx5 setup from another machine. Also trigger for adjacent symptoms like 'gnome-terminal 能用中文但 ghostty 不行' or 'ghostty 输入候选窗不弹'. The skill provides a 5-pitfall checklist (GTK4-XIM removal, snap classic GTK module loading, --gtk-single-instance zombie, desktop-entry priority shadowing, env propagation paths) and a /proc-based diagnostic protocol that pinpoints the exact failure layer in under 60 seconds. Prefer this skill over generic 'install fcitx5' guidance whenever the user already has fcitx5 working in other apps (e.g., gnome-terminal, browser) but ghostty is the holdout — that pattern is the high-value case

写出 reviewer 一页之内就买账、相信 story 合理 / 必要 / 有创新的英文 paper Introduction。方法论来自三源融合： (1) Western 写作大师 (Widom / SPJ / Black / Ernst / Tokekar / Williams) 的 5-段脚手架与 Goal-Problem-Solution 节律； (2) 中文 PhD 场景 (Sida Peng / MLNLP / LARG pre-writing form) 的可审计段落规则； (3) 10 篇 famous VLA / robot-learning 论文 (OpenVLA / π0 / π0.5 / π*0.6 / π0.7 / MemoryVLA / RoboMME / MEM / Fast-WAM / DiT4DiT) Introduction 全文逐句对照后提炼出的语言级共性。Skill 提供：6 段黄金骨架、4 类开篇 hook archetype、6 类段落 archetype 的句模与多篇真实 paper 实例、Black-Mad-Libs 风格的 Introduction 整段填空模板、动笔前 13 问表、25 项自检清单、ESL 中文母语者高频雷区列表。 Use PROACTIVELY whenever the user says "写 intro", "写 introduction", "改 intro", "improve introduction", "introduction 没写好", "story 不顺", "怎么开篇", "intro 节奏", "intro 怎么 sell", "rewrite intro", "introduction draft", "怎么 hook reviewer", "怎么写 motivation", "motivation 不够强", "introduction polish", "paper 第一节怎么写", "顶会 paper 第一页怎么写", "审稿人不买账", "我的 intro 太平", "intro 没有 story", "improve the framing", "intro 改 framing", "怎么写 contribution", "contribution 不亮", "怎么写 motivation 段", "怎么

IEEE RA-L 论文审稿助手。输入一篇待审稿 PDF，自动完成：初读论文提取关键信息 → 多源文献检索 (Semantic Scholar + WebSearch/arXiv + vec-db) → 并行 agent 深读相关论文 → 带着领域知识精读 待审稿论文 → 输出完整的 RA-L 审稿意见（含评分、推荐、双语评审意见）。审稿风格追求独到犀利， 不千篇一律。Use PROACTIVELY whenever the user asks to review a paper for RA-L, IEEE Robotics and Automation Letters, or says "审稿", "review this paper", "帮我审稿", "写review", "RA-L review", "审一下这篇", "peer review", or provides a PDF and mentions reviewing. Also trigger when user mentions PaperCept, reviewer form, or review deadline.

Create a structured literature survey database in Notion with paper metadata (title, arXiv link, venue, category, relevance). Use PROACTIVELY whenever the user wants to organize research papers, build a literature table, create a related-work database, or track surveyed papers in Notion. Also trigger on phrases like "paper table", "literature database", "把论文放到notion", "创建文献表格", "paper list to notion", "survey table", or any request to push paper references into Notion.

Implements Manus-style file-based planning to organize and track progress on complex tasks. Creates task_plan.md, findings.md, and progress.md. Use when asked to plan out, break down, or organize a multi-step project, research task, or any work requiring >5 tool calls. Supports automatic session recovery after /clear.

Help users work with the UCR HPCC (High Performance Computing Center) cluster. Provides commands for connecting, submitting jobs, managing software, and handling data storage.

Cross-platform information gathering from Xiaohongshu, Bilibili, Zhihu, and X/Twitter. Trigger when user wants to search, collect, or gather information across social platforms, says '全网搜索', '多平台搜索', '从各个平台搜', '搜集信息', 'search all platforms', '帮我搜一下', '调研一下社交媒体上的讨论', or mentions wanting opinions/discussions from Chinese social media and Twitter. Also trigger when user gives a topic and expects comprehensive coverage across platforms.

Automated experiment workflow for code + data experiments with structured planning, execution, and reporting. Trigger when user wants to run a full experiment cycle, says "开始实验", "跑一下验证", "实验流程", "setup experiment", "run experiment", "验证一下", "搭建实验", or needs workspace creation + iterative experiment loop + final reporting.