Hunt vector / embedding weaknesses (OWASP LLM08:2025) — adversarial inputs against the RAG / similarity layer that cause cross-tenant leak, embedding-inversion privacy loss, semantic confusion, and retriever-driven prompt injection.

Hunt LLM unbounded consumption (OWASP LLM10:2025) — denial-of-wallet and denial-of-service against LLM endpoints via unrestricted prompt size, runaway tool loops, expensive model selection, and unauthenticated fan-out.

Hunt SQL injection (CWE-89) via source-level taint tracking. Covers string concat, format-string, ORM raw queries, second-order injection, and NoSQL injection in MongoDB/DynamoDB.

飞书 IM 消息读取工具使用指南，覆盖会话消息获取、话题回复读取、跨会话消息搜索、图片/文件资源下载。 **当以下情况时使用此 Skill**: (1) 需要获取群聊或单聊的历史消息 (2) 需要读取话题（thread）内的回复消息 (3) 需要跨会话搜索消息（按关键词、发送者、时间等条件） (4) 消息中包含图片、文件、音频、视频，需要下载 (5) 用户提到"聊天记录"、"消息"、"群里说了什么"、"话题回复"、"搜索消息"、"图片"、"文件下载" (6) 需要按时间范围过滤消息、分页获取更多消息

Use when the user asks about content in their OpenKB knowledge base — research topics, concepts compiled from their documents, cross-document synthesis — or mentions `openkb`, an `.openkb/` directory, or a `wiki/` tree generated by openkb. The user may invoke you from any working directory; the active KB resolves via `openkb status`. Do NOT use for arbitrary Markdown directories, Obsidian vaults, or documentation sites not built by openkb.

Hunt LLM excessive agency (OWASP LLM06:2025) — agentic systems granted too many tools, too broad permissions per tool, or unsupervised authority to act on the user / business behalf, producing financial loss, data loss, or destructive operations from a single bad token.

Hunt LLM training-data and model poisoning (OWASP LLM04:2025) — adversarial inputs that bias future model behaviour through fine-tuning, RLHF, or continuous-learning loops.

Discover, analyze, and prioritize keywords for SEO and GEO content strategies from a seed keyword or niche. Identifies high-value opportunities based on search volume, competition, intent, and business relevance. Generates topic clusters. For building a dated editorial calendar from your own Search Console data, use `content-planner` instead — this skill is seed-driven keyword discovery, not calendar scheduling. Use when asked to "find keywords", "keyword research", "keyword analysis", "search volume", "keyword difficulty", "content ideas from a seed", or any keyword discovery task.

Submission-grade Nature/high-impact journal figure workflow for Python or R. Use whenever the user asks to create, revise, audit, or polish manuscript figures, multi-panel scientific plots, or journal-ready SVG/PDF/TIFF outputs, especially for Nature-family or other high-impact journals. Before plotting, define the figure's conclusion, evidence logic, export needs, and review risks. If the user has not chosen Python or R, ask "Python or R?" and stop. Use only the selected backend for figure generation, previewing, exporting, and QA. Supports matplotlib/seaborn and ggplot2/patchwork/ComplexHeatmap. Not for dashboards or Illustrator/Figma-first infographics.

Generate or edit images via Nano Banana image models. Triggers on "generate image", "image generation", "nano banana", "edit image", "nano banana pro", "nano banana 2"

Hunt OS command injection (CWE-78) — user input reaching shell, exec, or system calls. Covers argument-array bypasses, path confusion, and template-string injection in modern frameworks.

Activate when code touches token management, credential resolution, git auth flows, GITHUB_APM_PAT, ADO_APM_PAT, AuthResolver, HostInfo, AuthContext, or any remote host authentication -- even if 'auth' isn't mentioned explicitly.

Use this skill to drive a batch of suspected bugs in microsoft/apm from raw issue list to mergeable PR queue. Fan out one triage subagent per issue (LEGIT / UNCLEAR / FIXED-AT-HEAD), gate every legit bug against PRINCIPLES.md via an apm-ceo strategic-alignment pass, cross-reference legit issues against open PRs, then open a fix PR (TDD + mutation-break gate) for greenfield bugs. Drive every PR -- community in-flight and own fix alike -- to mergeable by composing the shepherd-driver skill: one driver per PR runs the review panel, folds non-blocking recommendations, pushes (preserving author), and watches CI to green. Re-probe mergeability and resolve conflicts via shepherd-driver. Maintain a plan.md ground-truth table as canonical state. Activate when the maintainer asks to triage issues, sweep the bug queue, shepherd bug-flagged issues, run a weekly community sweep, or drive in-flight community PRs to merge -- even if "shepherd" or "batch" is not named.

Use this skill to run a multi-persona expert advisory review on a labelled pull request in microsoft/apm. The panel fans out to five mandatory specialists plus a test-coverage specialist (active on every PR that touches src/) plus three conditional specialists (auth, doc-writer, performance-expert), all running in their own agent threads, and a CEO synthesizer. The orchestrator is the sole writer to the PR: ONE recommendation comment, no verdict labels, no merge gating. The panel is advisory -- it surfaces findings, prioritizes follow-ups, and renders a ship-recommendation that the maintainer and author weigh. Activate when a non-trivial PR needs a cross-cutting recommendation (architecture, CLI logging, DevX UX, supply-chain security, growth/positioning, optionally auth, docs, perf, and test coverage, with CEO arbitration).

Use this skill to drive any open microsoft/apm issue (bug, feature, docs, refactor, perf) from raw intake to a mergeable PR with triage as the central, paramount gate. Run the apm-triage-panel rubric per issue first, then present ONE consolidated triage review for the whole batch and escalate to the maintainer BY DEFAULT on any doubt (needs-design, decline, duplicate, defer, auto-handle, breaking- change, auth/security/governance surface, low arbiter confidence, unbounded scope, or a missing brief); only auto-implement clear, bounded, high-confidence accepts the maintainer approved. Then drive each accepted PR to mergeability batch-bug- shepherd style via the shepherd-driver loop: fold copilot + panel follow-ups by default, watch CI green, iterate under a bounded cap. Invoke MANUALLY, in-session, on an issue list or queue -- never by label or event. Activate when the maintainer asks to auto-tackle the issue queue, clear the backlog to PRs, or run issues to merge -- even if "autopilot" is not named.

Pre-PR validation — working tree status, format, lint, TypeScript check, and protected file check. Workspace-aware: detects which of backend/, frontend/, services/apps/*, services/libs/* were changed and runs only the relevant scripts. Use before submitting any PR, to check if code is ready, or to validate a branch before review.

Use this skill when implementing a new feature or fixing a bug. Write or update tests before marking the task done. Never consider code complete without verifying it works through automated tests.

Use this skill when drafting emails, Slack messages, announcements, or any external/internal communication. Apply professional structure and appropriate tone before writing any message.

Use this skill when a tool call, command, or API request fails. Diagnose the root cause systematically before retrying or changing approach. Do not retry the same failing call without first understanding why it failed.

花叔Design（Huashu-Design）——用HTML做高保真原型、交互Demo、幻灯片、动画、设计变体探索、设计方向顾问与专家评审。HTML是工具不是媒介，根据任务embody UX设计师/动画师/幻灯片设计师/原型师，避免web design tropes。触发词：做原型、设计Demo、交互原型、HTML演示、动画Demo、设计变体、hi-fi设计、UI mockup、prototype、设计探索、做个HTML页面、app原型、iOS原型、导出MP4/GIF、60fps视频、设计风格、配色方案、视觉风格、评审、review this design、带解说动画、voiceover、narration、TTS+动画。主干能力：Junior Designer工作流、反AI slop清单、React+Babel最佳实践、Tweaks变体切换、Speaker Notes、Starter Components、App原型交互守则、Playwright验证、HTML动画到MP4/GIF导出、带解说长动画pipeline。需求模糊时进入设计方向顾问模式：推荐差异化方向、展示showcase、并行生成视觉Demo供选择。交付后可做专家级5维度评审与修复清单。