Claude Skills
The open format is called Agent Skills and works in Claude Code, Codex, Cursor and other agents — most people know it as Claude Skills.
Every Agent Skill we could find on GitHub, deduplicated by content. 79 600 files from 1 763 authors, of which 61 947 are unique — the rest is the same skill repackaged into someone else's repository. For each one: what it weighs in tokens, whether it ships runnable scripts, and which MCP servers it needs.
> Reprice swap transaction legs under ASC 845 (Nonmonetary Transactions) to ensure clearing accounts apply equal-value exchange to swap legs, run setBatchUseCounterpartyFiatValue, zero out a clearing account, apply ASC 845, fix swap pricing mismatches, ensure no gain/loss on swaps, or close a month where a swaps/trade clearing account has a residual balance. Also trigger when the user mentions "counterparty fiat value", "swap repricing", "clearing account net to zero", or "equal value exchange".
> Guide users through creating ERP accounting rules for crypto organizations on the Tres Finance platform. Maps every transaction type to the correct chart of accounts entries using a layered rule engine.
> Trigger on-chain data collection (a "Commit") in Tres Finance for wallets that have already been onboarded. Use this skill whenever the user wants to collect data, pull balances, sync wallets, refresh on-chain data, run a commit, trigger a commit, or "collect" anything in Tres. This is the second step of the Tres onboarding flow — it sits between wallet upload (tres-wallets-upload) and balance validation (tres-asset-balance-validation). Always trigger this skill for any request that mentions "commit", "collect", "data collection", "pull data", "fetch on-chain data", "sync wallets", "refresh balances", or anytime the user just finished uploading wallets and is ready to bring in their on-chain data.
Extract all 3rd-party (non-owned) addresses from a TRES Finance environment and export them as an XLSX workbook to help users build their contacts list. Use this skill whenever the user wants to export, list, or identify 3rd-party addresses, counterparties, or external addresses from their transactions — even if they don't say 'contacts'. Also trigger when the user asks to prepare a contacts import file, find unknown addresses, or build an address book from transaction history. Do NOT trigger for viewing existing contacts or searching the address book — only for extracting NEW 3rd-party addresses from transaction data.
Import contacts (address book entries) into TRES Finance from a CSV or XLSX file. Use this skill whenever the user wants to import, upload, or bulk-add contacts, address labels, or address book entries into TRES — whether from a file they filled in after using the export-3rd-party-contacts skill, or from any CSV/XLSX that maps addresses to names. Trigger phrases include: 'import contacts', 'upload contacts', 'add contacts from file', 'import address book', 'load contacts into TRES', 'bulk label addresses', 'import the contacts file', 'upload the filled contacts sheet'. Also trigger when the user says something like 'I filled in the contacts file, now import it' or 'here is my contacts spreadsheet, please upload it to TRES'. Do NOT trigger for exporting or extracting unidentified addresses — that's the export-3rd-party-contacts skill. Do NOT trigger for viewing or searching the existing address book.
> Match TRES ledger transactions to ERP invoices/bills (AP/AR) and optionally sync them to the connected ERP (Xero, QuickBooks Online, NetSuite). Trigger this skill whenever the user wants to match, link, close, reconcile, or sync an invoice or bill against a blockchain transaction — even if they don't say "skill" or "close invoice INV-123", "close bill 9988", "link this tx to a bill", "pay this invoice from this transaction", "set up this tx as AP", "set up this tx as AR", "sync this transaction as AP/AR", "match AP/AR", "find the invoice/bill for this hash", "what bill does this tx pay". Trigger ONLY for explicit AP/AR matching/closing intent — do NOT trigger for general transaction explanations (use tres-tx-story), for ingesting an explorer link into the ledger (use tres-explorer-tx-to-ledger), or for ERP connection setup itself (use tres-settings-management).
Recommend the right TRES Finance report for any user question. Trigger this skill whenever a user asks which report to use, what report contains certain data, how to get specific information out of TRES, or compares two reports. Also trigger when the user describes a goal (e.g. "month-end close", "auditor needs proof", "tax filing", "reconciliation") without naming a specific report. Trigger phrases include "which report", "what report", "how do I export", "where can I find", "I need a report for", "what's the difference between", "best report for", "auditor asked for", or any question about getting data out of TRES Finance.
> Query, display, and resolve reconciliation gaps from the TRES Finance MCP connector. Trigger this skill ONLY when the user explicitly requests to view or close reconciliation for today", "plug the reconciliation gaps for ETH", "fix the gaps for USDC and BTC". Do NOT trigger for general balance queries, transaction history, or ledger browsing. fetching gap data, enriching with fiat and on-chain balances, rendering an HTML dashboard where each gap row has copy-prompt buttons (Plug / Auto-fill) — clicking copies a ready-to-paste Claude prompt for that gap. The user pastes it in the chat and Claude executes the action. Ends with a data collect once the user is done.
> Build a TRES Finance dashboard ledger URL (Transactions tab) with a precise filter set — date range, wallets, assets, tags, status, amount, and more — so it can be shared with a customer or reused. Use whenever the user asks for a "ledger link", "dashboard link", "filter URL", or "send <customer> a link to <some filtered view>". Scoped to the Transactions tab only (not Accounting, Roll Forward, Pivot Tables, Cost Basis, or Trial Balances).
> End-to-end onboarding of a new entity, customer, or account into TRES Finance. validation, reconciliation, cost basis, export unidentified addresses, import contacts, and rollup rules. Trigger ONLY when the user explicitly wants to onboard a new entity, customer, or account — e.g. "onboard a new entity", "onboard new customer", "start onboarding for [name]", "set up a new account in TRES", "run the full onboarding flow", "I need to onboard [company name]", "new entity onboarding". Do NOT trigger for individual tasks — those have their own dedicated skills. If the user asks to "add wallets" without mentioning onboarding, use tres-wallets-upload. "Validate balances" alone is tres-asset-balance-validation. "Run a commit" alone is tres-data-collection-commit. Only trigger when the user wants the full pipeline.
Analyze any TRES Finance report XLSX and produce an automatic findings summary. Trigger this skill whenever a user uploads a .xlsx file that came from TRES Finance and asks to "analyze", "summarize", "review", "check", "audit", "look at", "what does this show", "any issues", "anything interesting", or "walk me through" the report. Also trigger when the user uploads an Excel file and mentions TRES, reconciliation, balances, transactions, cost basis, roll forward, or ERP sync. The skill identifies the report type from the file structure, runs targeted analysis based on report-specific logic, and produces a clear findings summary highlighting anomalies, key metrics, and action items.
> Submit feedback about the TRES Claude plugin — bug reports, feature requests, skill improvements, new skill ideas, MCP issues, workflow friction, or positive feedback about what's working well. Use this skill whenever the user wants to share any kind of feedback "report a bug", "request a feature", "suggest an improvement", "I have an idea", "something is broken", "the skill isn't working", "I love this skill", "this worked great", "the MCP is returning wrong data", "the flow is confusing", "can you fix the skill", "file a skill request", "give feedback", "this could be better". Also trigger proactively when the conversation appears to be wrapping up — e.g. the user says "thanks", "that's all", "looks good", "we're done", "nothing else", or any other signal that the task is complete. In that case, offer a brief feedback prompt instead of launching the full guided flow. Do NOT trigger for general questions about how a skill works — only when the user wants to submit feedback or the conversation is ending.
Create (generate) any TRES Finance report end-to-end via the tres-mcp MCP server and GraphQL — trigger the export, verify a report row was actually created (guarding against silent failures), poll until done, and return the download link. Trigger this skill on ANY report creation request, including phrases like "create a report", "generate a report", "export", "run a report", "I need a Transaction Ledger / Balances / Reconciliation / Cost Basis / Roll Forward / Staking / Audit report", "pull the data for", "download a report for", "give me a CSV/XLSX of", or whenever the user wants TRES to produce a report file. Use together with tres-report-advisor when the user is unsure which report they need.
Manage Organization Settings and Platform Settings via the TRES MCP GraphQL API. Use when users ask about org settings, platform settings, configuration, feature flags, enable/disable platforms, balance diff, commit strategy, cost basis, ERP, pricing, sync boundaries, or any setting read/write operation. Trigger phrases include "get settings", "show settings", "update settings", "change settings", "enable platform", "disable platform", "balance diff", "commit strategy", "cost basis strategy", "set min sync date", "configure", "turn on", "turn off".
> Compute the EXACT sub-transaction impact of a rollup rule in TRES Finance — every filter applied — and flag genuine problems (zero matches, overlap, config mistakes) of a rule someone submitted or of existing PENDING rules. Trigger this skill when the user says "review this rollup rule", "what would this rollup rule actually match", "preview the rollup impact", "how many transactions would this rule catch", "is this rollup rule valid", "validate my rollup rule", "check my pending rollup rules", or "review the rollup requests". Do NOT use this to design or propose new rollups — that is tres-rollup-rules. This skill only measures impact and flags issues; it never executes mutations.
The exact column header names required by the TRES Finance bulk transaction CSV upload. Use whenever building, fixing, or validating the header row of a TRES manual transaction or sub-transaction upload CSV — especially after an upload fails with \"Missing required column(s): ...\". Trigger phrases include: 'fix my CSV headers', 'bulk transaction upload failed', 'missing required columns', 'validate TRES upload CSV', 'transaction CSV template headers'. This skill is ONLY about the header naming convention; it does not cover row values or upload workflow. Do NOT trigger for contacts import CSV (use tres-import-contacts), wallet upload CSV (use tres-wallets-upload), or general CSV parsing unrelated to bulk TX upload headers.
> Analyze a blockchain transaction by its hash using TRES Finance MCP and generate a visual flow diagram showing all asset movements, followed by a plain-language explanation of what happened. Use this skill whenever the user provides a transaction hash and asks to "analyze", "explain", "visualize", "diagram", "show me what happened in", or "tell the story of" a transaction. Also trigger when the user pastes a tx hash (long hex string starting with 0x) and asks anything about it. Always use this skill for transaction hash analysis — do not attempt to answer from memory alone.
> Analyze an organization's sub-transaction patterns in TRES Finance, identify groups of repetitive transactions, and propose data-driven rollup rules that consolidate them into one aggregated entry per interval (daily or monthly) for cleaner ERP integration. Also lists and deletes rollup rules. Trigger this skill whenever the user asks about rollup rules, transaction aggregation, consolidating transactions, reducing transaction noise, rolling up gas fees, rolling up staking rewards, rolling up micro-transactions, or collapsing many small ledger entries into fewer summary entries. Also trigger on "create a rollup rule", "suggest rollup rules", "what should I roll up", "show my rollup rules", "delete a rollup rule", "aggregate transactions", "too many transactions", "reduce transaction count", "roll up fees", "daily rollup", "monthly rollup". If a wallet has too many transactions, high-frequency activity, or noisy ledger entries, proactively suggest rollup rules. For computing the EXACT impact of a single already-specified rule or reviewing submitted/pending rules, use the tres-rollup-review skill instead.
> Upload and onboard multiple on-chain wallets or exchange accounts into Tres Finance. Use this skill whenever the user wants to add wallets, onboard wallets, import wallets, upload wallet addresses, ingest wallets, or connect exchanges into Tres — whether from a file (CSV/Excel) or by typing them manually. This skill guides Claude through the full HTML table → exchange credential collection (if exchange) → user confirmation → batch creation via the TRES MCP API. Always trigger this skill for wallet onboarding, wallet data ingestion, or any request to "add wallets to Tres".
Install and configure the anti-slop Oxlint plugin in a local TypeScript or JavaScript repository. Use whenever a user asks to add anti-slop lint rules, copy the anti-slop plugin, configure opinionated Oxlint rules, or migrate an existing local anti-slop setup.
The user asks a question about a video that was already watched or indexed — "what did they say about X", "what error code appears", "what happens at 2:30", "does the video show Y". Use this to answer from the persistent index with timestamped evidence and a confidence score instead of re-watching or guessing.
The user asks about videos watched in the past or across sessions — "have we watched anything about X", "which video showed that error", "what did that meeting decide", "search my videos", or a question that spans several videos. Use this to search and answer from the persistent cross-video index instead of saying you don't remember.
The user built or changed something visual — a UI, an animation, a game, a generated video — and wants it verified, or asks "why does my UI look wrong", "check that the fix actually worked", "does the animation glitch". Use this to record the running thing, critique the recording against plain-language pass criteria, and iterate until it passes with before/after proof.
The user wants to share or hand off what was found in a video — "send this analysis to my team", "make a page I can share", "export the findings", "give me something I can attach to the ticket". Use this to render a self-contained offline HTML viewer page with the frames, transcript, and every cached answer with its evidence.
The user corrected an answer about a video — "no, it actually says X", "that's the wrong timestamp", "you misread the error code" — or asks why a video answer was wrong. Use this to record the correction as a lesson so future answers on similar questions improve, and to show what the system has learned and saved.
The user wants to connect an LLM or vision provider, already has an API key, asks "can I use OpenAI/Anthropic/Gemini/OpenRouter", wants local Ollama, or needs different cheap and strong models. Use this to configure provider-neutral visual understanding without tying Watch Skill to one agent or model vendor.
A Watch Skill command failed, video download broke, ffmpeg or yt-dlp is missing, a provider returned an error, local vision stopped, or the user asks "why can't it watch this" or "repair my setup". Use this to diagnose and apply the engine's structured fix before retrying the original operation.
The user wants structure pulled out of a watched video — "make chapters for this video", "where does the bug appear in this recording", "turn this screen recording into a bug report", "how strong is my intro/hook". Use this for deterministic extraction from the index — chapters with timestamps, a fileable bug report with the exact frame, or a scored hook analysis.
Watch any video (URL, stream, or local path) via Watch Skill. Downloads, extracts scene-aware deduped frames, OCRs them, transcribes (captions first, then local Whisper — offline by default), indexes everything, and hands the result to the agent. Follow-up questions are answered from the persistent index without re-processing.
The user shared a video URL, a YouTube/TikTok/stream link, a local video file, a screen recording, a meeting recording, or a playlist/folder of videos — "watch this", "summarize this video", "what's in this recording". Use this to actually watch the video — download, extract frames, OCR, transcribe, and index it — instead of guessing from the title or asking the user to describe it.
在创作者明确确认后，执行短剧项目的图片、视频、TTS/配音或时间线音乐生产任务，并把结果与精简运行记录落回项目。用户说“生成这张图/这段视频/这句配音/这段配乐”“开始跑图/跑视频/合成语音/生成音乐”“把已确认提示词送去生产”，或要求批量执行已确认媒体任务时使用；不负责创作提示词、镜头、台词、歌词或声音身份，也绝不把预览、继续、预算说明或既有接受状态当作本次付费生产确认。
把已经生产出来的逐镜素材按分镜装配成短剧成片：定每镜入出点、镜序、台词完整性、字幕、响度与交付规格，写成 `剧集/<EP>/剪辑单.md` 并渲染出成片。用户说“把这些镜头剪成片”“做成片/装配/套剪”“定入出点”“加字幕”“统一响度”“成片太长/太碎”“导出可继续精修的工程”时使用；不生成新素材，不改剧本、分镜或视频提示词的语义。
审查短剧项目中的原著分析、故事、剧本、视觉设定、图片提示词、分镜、冻结关键帧、视频提示词和已有媒体。用户提出“审稿/检查剧本”“检查资产或连续性”“检查图片/视频提示词”“检查原著分析”“审查模板感”“根据生产观察做项目校准”时使用；只写审查问题、结论和修订要求，不代替 owner 修改来源文件。
为短剧人物、造型、地点、道具和状态编写或修改可直接复制的图片提示词 Markdown。用户提到角色设定图、三视图、参考图、场景板、道具图、风格帧、Look Development、状态变体或局部编辑提示词时使用；不生成图片，也不调用供应商。
把中文短剧剧本和视觉设定写成有戏剧职责、连续性边界与冻结关键帧提示词的分镜 Markdown。用户提出“拆分镜/设计镜头/做镜头表”“场次视觉计划/调度故事板”“比较导演方案”“写首帧/关键帧提示词”或检查轴线、站位、视线、持物连续性时使用；不生成媒体。
从短剧剧本拆出人物/造型、地点/视图、道具/状态和跨场连续性，写成创作者可读的视觉设定。用户说“拆角色/场景/道具”“做资产设定”“判断复用还是新变体”“更新造型/道具状态”，或拿现成剧本直接做视觉资产准备时使用；不写图片提示词，不生成媒体。
编写或修订可拍摄的中文短剧、漫剧单集 Markdown 剧本，也负责保留作者原文地规范化现成剧本。用户提出“写/改一集短剧”“把大纲写成剧本”“优化场景/对白”“去模板感”“去 AI 味润色”“续写下一集”或提供剧本要求进入后续制作时使用；不负责资产、分镜、媒体提示词或终审。
把短剧分镜和冻结关键帧写成可直接复制的视频提示词 Markdown，也可按用户要求写时间线配乐/主题曲意图。用户提到文生/图生视频动作、人物表演、运镜、口型、环境运动、镜头时长、起止状态、把分镜转成视频提示词或写配乐提示词时使用；不生成媒体、不创作歌词、不改分镜边界。
基于文件系统初始化和继续短剧或漫剧项目，提供 creator-first 五文档路由、本地 Dashboard、制作形态与 Look Development 决策。用户提出“创建/继续短剧项目”“看进度/下一步”“做 Look Development”“打开 dashboard/短剧创作台”“导出制作资料”，或任务跨多个创作阶段时使用；明确的写作、资产、提示词、分镜、剪辑或审查请求由对应子 skill 直接处理。
Design game concepts from a novel. From SOURCE_BIBLE and PRODUCT_BRIEF, generate three genuinely different directions on the dimensions still unlocked, then pick the most worthwhile playable prototype using hard vetoes and explicit trade-offs. Use for what game should this novel become, compare game concepts, choose a game direction for this book, decide whether this novel should become an interactive story. 小说游戏概念设计。根据 SOURCE_BIBLE 与 PRODUCT_BRIEF 在未锁定维度上生成三个真正不同的方案，用硬否决和关键取舍选择最值得做的可玩原型。用于判断小说适合做成什么游戏、该不该做成互动叙事、比较游戏方案等需求。
Verify a game with evidence on its selected target runtime. Launch the actual build and prove real rendering, input, the core loop, at least one designed outcome, restart, and explicit limitations without dressing subjective fun up as a certain verdict. Use for test a generated game, QA a game build, check whether the game is fully playable, or verify the build. 游戏证据化质量验证。在选定的目标运行环境中启动实际构建，证明真实渲染、输入、核心循环、至少一个设计结果、重开和明确限制，不把主观趣味包装成确定性结论。用于测试生成游戏、检查游戏能否完整游玩或验证构建。
Direct game art and creative vision. Turn GAME_DESIGN into a production-level ART_DIRECTION defining a recognizable visual style, camera and composition, world and character grammar, functional colour/light/material, HUD feedback, motion and transition specs, audio direction, and a signature moment for every screen and mode. Use for what should the game look like, set the art direction, define the visual style. 游戏美术与创意方向。把 GAME_DESIGN 转成策划级 ART_DIRECTION，定义可辨识视觉风格、镜头构图、世界与角色语法、功能性色光材质、界面反馈、运动与转场规格、声音方向和每个界面/模式的签名游戏时刻。用于判断游戏应该长什么样、制定游戏美术方向等需求。
Design game experience, systems, and levels. Converge the chosen concept into one GAME_DESIGN defining the player promise, core loop, how the world responds, the systems actually needed, level pacing, feedback, failure, and a fully playable prototype. Use for design the game world, deepen the gameplay and levels, write the game design document, design the branching story structure, scenes, dialogue and consequences. 游戏体验、系统与关卡策划。把选定概念收束为一份 GAME_DESIGN，定义玩家承诺、核心循环、世界响应、必要系统、关卡节奏、反馈、失败和可完整游玩的原型。用于设计游戏世界、深化玩法和关卡、设计互动叙事的场景对白与分支因果等需求。
Build a risk-matched whitebox or the approved production game for its target runtime. Turn GAME_DESIGN, and ART_DIRECTION when production begins, into a minimal BUILD_BRIEF and a runnable candidate that can be iterated with replayable evidence. Use for prototype the riskiest design question, implement the approved game design, or turn this design into a running game. 游戏构建执行。先按最大风险做白盒，或在正式生产时把批准后的 GAME_DESIGN 与 ART_DIRECTION 压缩成最小 BUILD_BRIEF，在目标运行环境中实现可运行候选并用可回放证据迭代。用于验证高风险设计问题、构建游戏原型或实现批准方案。
Deconstruct a novel for game adaptation. Compress a raw novel, deconstruction library, or writing project into a SOURCE_BIBLE with cited textual evidence, extracting world rules, player verbs, spaces, character will, systems, and visual anchors — without inventing a genre yet. Use for gameable book analysis, analyze a novel for game adaptation, extract game systems from a book. 游戏化拆书。把原始小说、拆文库或写作工程压缩成有必要原文证据的 SOURCE_BIBLE，提取世界规则、玩家动词、空间、角色意志、系统和视觉锚点，不提前发明游戏类型。用于游戏化拆书、分析小说的游戏改编基础等需求。
Turn a novel into a fully playable game on the selected target platform. Orchestrates the whole adaptation pipeline — requirements intake, gameable deconstruction, concept selection, world and visual design, target-runtime build, and evidence-based QA — for a novel in any language. Use for novel to game, story to game, book to game, adapt this novel into a game, turn this book into a playable prototype, make an interactive story or text adventure from this novel. NovelToGame 总入口。把任意语言的原始小说、拆文库或 oh-story 写作工程转成有原著依据、可在目标平台完整游玩的游戏，编排游戏化拆解、概念选择、游戏与视觉设计、目标运行环境构建和证据化质量验证。用于小说转游戏、把这本书做成游戏、把小说改成互动小说 / 互动叙事 / 文字冒险游戏等需求。
网文去AI味。检测并清除文本中的AI写作痕迹，让文字回归自然、非模板化。触发方式：/story-deslop、/去AI味、「去AI味」「这篇太AI了」「网文去AI味」。
长篇网文拆文。深度拆解爆款长篇小说的黄金三章、人设架构、爽点设计、节奏控制。单一深度拆解管道：跑完黄金三章（Stage 1）后产出快速预览报告并询问是否继续全量拆解，确认后从 Stage 2 续跑逐章摘要、聚合分析、设定关系、汇总报告，全程产物落盘 拆文库/{书名}/。触发方式：/story-long-analyze、/长篇拆文、「帮我拆这本书」「拆这本书」「分析黄金三章」「深度拆解」「完整拆解」「系统拆解」或提供小说文本文件路径——全部进入同一管道。
逆向导入已有小说。将已写好的小说（半成品或完本）反向解析为标准项目目录结构，兼容 story-long-write / story-short-write 后续写作流程；内部复用 story-long-analyze / story-short-analyze 的拆解管道，按篇幅自动分流。触发方式：/story-import、「导入小说」「反向解析」「导入」「把我的书导进来」。
长篇网文写作。从大纲到正文，辅助长篇网络小说的创作，包括世界观、人物、情节线管理。触发方式：/story-long-write、/写长篇、「帮我开书」「写大纲」「日更」「续写」「继续写」「修改第X章」「回炉」「重写第X章」。
多视角对抗式审查。full/lean 模式在已部署 reviewer agents 时并行 spawn；缺失/异常 agents 或 spawn 失败时自动降级 solo，参考文件不可读时使用内置 rubric fallback。触发方式：/story-review、/审查、「审查一下」「帮我审一下」。
短篇网文拆文。拆解爆款短篇小说（番茄短篇 / 故事会 / 知乎盐选 / 追妻 / 世情 / 重生 / 虐渣等通俗题材）的故事核、结构、情感线、反转设计、写作手法、共鸣层次。单一全量拆解管道：跑完 Stage 2-6 产出完整拆文报告，落盘到 拆文库/{书名}/，下游 story-short-write 同时读拆文报告 + 情节节点 + 写作手法 + 原文 + _meta.json 写下一篇。触发方式：/story-short-analyze、/短篇拆文、「拆短篇」「拆这篇短文」「短篇拆文」「精细拆解短篇」「8000 字短篇拆解」「番茄短篇拆文」「故事会拆解」「盐言故事拆解」「分析这篇短篇」——均进入同一管道。
短篇网文写作。辅助短篇小说创作，从构思到成稿，聚焦情绪拉扯与节奏把控。触发方式：/story-short-write、/写短篇、「帮我写一篇短篇」「写个盐言故事」。
网络小说工具箱主入口。根据用户需求自动路由到对应 skill，并可管理作者习惯、启动本地 Dashboard。触发方式：/story、$story、/story dashboard、/网文、「我想写小说」「记住我的写作习惯」「打开工作台」「检查更新」。
网文写作工具集基础设施部署与检查。为 Claude Code / OpenCode / Codex / Google Antigravity / ZCode / OpenClaw / Reasonix 提供内置适配；Web AI / 通用 Agent 可走 skills + AGENTS.md 文件模式。触发方式：/story-setup、$story-setup、「准备写书」「帮我搭一下环境」「配置写作项目」「检查写作环境」。
Use when red-teaming an agentic AI / LLM application — indirect & zero-click prompt injection, MCP tool poisoning, persistent memory poisoning, excessive-agency tool abuse, multi-turn jailbreaks, PyRIT/Garak/Promptfoo harnesses
Use when attacking or auditing a CI/CD pipeline or software supply chain — pwn requests, poisoned pipeline execution, compromised/mutable-tag actions, dependency confusion, registry worms, runner backdoors, OIDC trust abuse, SLSA/provenance
Use when attacking an AI/ML system or model — prompt injection & jailbreaks (Crescendo, Skeleton Key, Best-of-N), RAG/vector poisoning, agentic/MCP exploitation (CVE-2025-54136), ML supply-chain RCE (pickle CVE-2025-32434), model extraction / membership inference / adversarial suffixes (GCG)
Use when attacking AWS/Azure/GCP cloud — IAM/identity privilege escalation, IMDS/metadata SSRF, Entra device-code & PRT theft, GCP impersonation chains, Kubernetes/container escape, IaC/CI-CD federation abuse
Claude Skills — questions
Answers built from the skills we actually parsed.