AX-skills

Build production-grade, phone-friendly terminal UIs (full-screen TUI dashboards) with zero terminal libraries — pure-core architecture, ANSI-safe layout math, line-diffed painting, SGR mouse + touch gestures, hit-testing, narrow-width responsive layouts, and a headless test harness. Distilled from the axfleet fleet-monitoring TUI. MUST trigger when the user says: "build a TUI", "terminal dashboard", "terminal UI", "full-screen CLI", "ncurses-style", "做个 TUI", "终端面板", "终端界面", "命令行 dashboard", "TUI 适配手机/Moshi", "mouse support in terminal", or asks to add tabs/mouse/touch/responsive behavior to an existing terminal program. Cross-platform: works on Claude Code, OpenAI Codex, OpenCode, and OpenClaw.

Activate BEFORE starting any /goal or self-driven autonomous loop. Suppresses end-of-turn status summaries, handoff prose, and "where to pick up next session" wrap-ups that prematurely signal completion. Companion hooks block Stop events when wrap-up patterns appear mid-loop, and block writes to HANDOFF/STATUS/NEXT-SESSION files until the loop is explicitly cleared. Invoke once at loop start with `start <predicate>`; clear on completion with `clear`. Composes with /goal — does not replace it.

Engineering workflow router — picks direct / light / full superpowers cycle based on RISK, not file count. AUTO-INVOKED for tasks touching 3+ files or creating new functionality, but defaults to LIGHT process unless the work has real architectural risk (concurrency, atomicity, protocol, schema). One entry point over the superpowers engine + audit/pr/sweep loops. Skip for bug fixes, typos, single-function changes, and questions. RUNS AUTONOMOUSLY by default — only pauses on Critical Decision Triggers (see Phase -1).

Design, audit, or improve durable context and memory systems for AI agents using layered observations, reflection, conservative promotion, progressive disclosure, and replayable boundary artifacts. Use when building long-running agent memory, deciding what belongs in sessions versus rules or skills, preventing context bloat, or turning repeated corrections and workflows into durable assets. Do not use it to copy another person's private axioms or automatically promote one-off observations.

Diagnose unclear, intermittent, cross-component, or high-impact bugs with evidence and falsifiable hypotheses. Use when the cause is not evident, a previous fix attempt failed, behavior differs across environments, or the failure crosses data or control-flow boundaries. If the cause is already local and obvious, use the fast path instead of turning the bug into an investigation ceremony.

Design or audit model capability, reasoning depth, agent roles, delegation boundaries, and review topology from the actual shape of work. Use when choosing between direct execution and delegation, defining a multi-agent workflow, assigning heterogeneous capability tiers, or correcting wasteful or shallow agent orchestration. Do not invoke it merely because agents are available or to protect the coordinator from reading important context.

Use when designing, auditing, or improving an AI-native workflow, agentic project, coding-agent setup, skill, pipeline, scheduled automation, or LLM-backed product where deterministic code and AI judgment interact.

卡兹克风格公众号长文写作skill——强口语化、叙事驱动、四层自检(L1硬性规则→L2风格→L3内容→L4活人感)的高质量内容产出。 适用场景：需要卡兹克式口语风格的公众号长文（"用卡兹克风格写"、"按khazix风格出稿"、"写一篇有活人感的文章"），或需要叙事驱动+文化升维的深度长文。 不适用：企业官号正式文案、多风格通用写作（用wechat-article-writer）、短内容（小红书/推特）。 触发词：卡兹克风格、khazix风格、活人感写作、写稿子、帮我写长文。

Use this skill when the user asks to "fix PR comments", "address review feedback", "get PR green", "fix CI and review comments", "loop until PR passes", or wants to autonomously respond to GitHub PR review feedback until all comments are resolved and CI is green. Max 5 rounds.

End-of-session knowledge cleanup with OCD-level rigor — reconciles project docs (CLAUDE.md, README.md, docs/) and agent memory against the code so nothing rots. 会话结束后对项目文档和记忆进行洁癖级审查与同步。MUST trigger when the user says: "sync up", "tidy up docs", "update memory", "clean up docs", "/sync", "/neat", "同步一下", "整理文档", "整理一下", "更新记忆", "梳理一下", "收尾", "这个阶段做完了", "新人能直接上手", or any phrase suggesting a dev milestone where knowledge needs reconciliation. Also trigger when the user reports stale docs, conflicting memories, or wants a clean handoff to teammates or other agents. Bare "整理" / "tidy" with prior dev context counts — do not under-trigger. Cross-platform: works on Claude Code, OpenAI Codex, OpenCode, and OpenClaw.

三合一东方占卜（小六壬 / 梅花易数 / 六爻）+ 赛博半仙人设。用户描述事情，半仙智能选起卦法，py 引擎出真实卦象，按 panpanmao 占卜知识库做半仙化解读。守"一事一占""医不问卦"等占卜规矩。

验证型调研：多Agent并行、交叉验证（Tier 1-4信源分级）、可追溯的调研工作流。产出决策备忘录，核心是判断claim的真伪和可信度。 适用场景：需要做决策的调研（"该选A还是B"、"X靠不靠谱"、"evaluate"、"帮我验证"）、技术评估、方案对比、行业可行性判断。 不适用：纯粹想了解一个东西的来龙去脉（用 narrative-research）、想要叙事风格的研究报告（用 narrative-research）。 触发词：深度调研、帮我调研、evaluate、该选哪个、靠不靠谱、对比一下、investigate、competitive analysis。

Generate images or image edits with Google Gemini / Nano Banana. Use for Gemini image, nano banana, multi-reference image edits, background swaps, character consistency, style transfer, or soft illustrated / watercolor / anime aesthetics; prefer gpt-image for readable text, large same-prompt batches, and sharp product/editorial defaults.

Generate images via OpenAI's GPT Image family (Azure deployment gpt-image-2, OpenAI direct fallback on rate-limit). PREFER this skill for: photorealistic scenes, product shots, editorial/brand imagery, UI mockups, and any image where embedded text MUST render correctly (posters with titles, infographics with labels, logos with wordmark). AVOID for: illustration, anime, watercolor, Chiikawa / 吉卜力 / hand-drawn aesthetics, reference-image editing, or character-consistency across scenes — use the separate nanobanana skill (Google Gemini) for those. When genuinely ambiguous (no style signal), ask once: "realistic / editorial (gpt-image) or illustration / anime (nanobanana)?" — do NOT silently default.

Parse complex PDFs (and images / DOCX / PPTX / XLSX) into LLM-ready Markdown + structured JSON with MinerU — tables, formulas, and reading order preserved. Use for research ingestion, RAG data prep, and academic papers / technical reports where layout fidelity matters. Triggers: "PDF解析", "PDF转Markdown", "提取PDF表格/公式", "parse this PDF", "MinerU". NOT for reading a few pages (just read them), or for merge/split/rotate/fill/OCR-in-place — use the `pdf` skill for those.

约束 AI 写好【中文网文长篇连载】的工业化工作流 skill —— 把"完全放手让 AI 写、4-5 章后逻辑/人设/世界观全崩"的失控过程，改造成"人类当导演、AI 当受约束子程序"的可稳定推进流水线。核心不是"写作 agent"，而是五件事：① 第0步人类锁定品类/平台/顶层契约（核心冲突·结局·主线锚点·底层规则）；② 把记忆拆成带 Canon 状态+按卷可见性的 typed 状态文档（世界观/人物/剧情线/伏笔台账/情绪债/滚动摘要），而不是一个会泄漏到正文的 memory.md；③ 用确定性流程每章编译"最小 context、指令与正文严格分离"的短 prompt（专治 AI 把约束写进正文）；④ 章节生产闭环（编译→生成→独立一致性校验→量化 rubric 审校→定向改稿→去 AI 味→状态增量回写），改稿≤3轮、reviewer 与 writer 独立调用防自欺；⑤ 反 AI 味分三桶处理（词句层全量复用文学清单·结构层改分层规则·网文专属轴爽点/钩子/毒点/品类适配）。明确划分 AI自动 / 人机协作 / 必须人类拍板 三档边界。 设计约束：model-agnostic（默认就能在 Codex/GPT 上跑，换 GLM 等国产文学性更强的模型是可选优化旋钮、非前置依赖）；诚实区分"工程能修的"（一致性崩坏/结构失控/prompt泄漏/流程纪律）与"模型天花板"（文学性/语感/叙事层 AI 味，工程压不到零）；反 overengineering（substrate 是 Markdown/YAML + skill 指令 + 极小可选校验脚本，不是要作者建一个带 SQLite/向量库/dashboard 的 app）。 适用：用 AI 写网络小说/网文连载，发现"放手写就崩、续写几章逻辑乱、人设漂移、伏笔忘填、AI 味重、节奏平没爽点"；想搭一套约束 AI 的写作流水线；要维护世界观/人物/剧情/时间线/伏笔 state；要做章节级生成+审校循环；要去 AI 味；要按品类/平台（起点/番茄/晋江/飞卢）适配。姊妹于 game-script-creation（那个是二游剧本，这个是网文长篇连载）。 触发词：写网文、网络小说、写小说、AI 写小说、长篇小说、连载、续写、小说崩了、AI 写小说逻辑乱、人设漂移、世界观崩、伏笔、章节生成、小说大纲、世界观设定、人物设定、爽点、黄金三章、章末钩子、毒点、去 AI

Scaffold a repository's agent-maintainability contract: an AGENTS.md architecture contract, an executable verification harness (boundary / giant-file / generated-clean checks), a single `make verify` / `pnpm verify` command, `.agent/` refactor ledgers, and optional Cursor rules. Stack-aware (frontend / backend / general). Use when the user says "scaffold AGENTS.md", "set up the agent contract", "add a verification harness", "make this repo agent-maintainable", "add architecture boundary checks", "wire up make verify", or starts/hardens a repo for long-horizon AI-agent maintenance. Implements the bundled agentic-repo-architecture blueprint (references/). Distinct from /init (which writes a codebase-documentation file): this writes the architecture CONTRACT + executable gates.

Use this skill to analyze a secular technology/industry trend, supply chain, company, or public investment thesis using a Serenity-style bottleneck research process. The skill converts a broad narrative into concrete constraints, evidence gates, disconfirmation tests, and next primary-source checks. It must not be used to copy trades, produce buy/sell signals, or treat social-media posts as proof.

Use when generating a custom desk-pet character for clawd-on-desk (or any APNG-based pet runtime) from either a text description or a real-pet photo + breed. Produces 10 lively, character-consistent state animations (idle / sleeping / working-typing / thinking / happy / notification / react-poke / error / collapse-sleep / wake), green-screen chroma-keyed APNGs, and a ready-to-install theme.json. Trigger phrases: 自定义桌面宠物, custom desk pet, generate pet animations, clawd theme, 给我做一只猫, pet-forge, 桌宠生成, custom cat sprites, AI pet animation pipeline.

Set up Cloudflare R2 object storage for a Vercel / Next.js app, or migrate off Vercel Blob, the de-risked way — proven write-toggle + host-routed-reads architecture plus the full pitfall list (wrangler can't mint S3 keys, Vercel CLI env corruption, next/image + CSP blocking R2 images, custom-domain CDN caching 404s, private-bucket PII reads, sharp/test gotchas). Use when the user says "use R2", "Cloudflare R2", "migrate off Vercel Blob", "object storage for Vercel", "S3-compatible storage", "image storage that works in mainland China", or sets up an R2 bucket for a server app.

This skill should be used when the user wants to "transcribe audio", "transcribe video", "speech to text", "convert audio to text", "transcription", or needs to extract text from audio/video files. Default engine is 妙记 (Volcano Lark Minutes ASR) for best speaker diarization; Gemini/OpenAI available as fallbacks. Supports files up to ~8.4 hours with speaker diarization.

「诚实高于善意」的深度自审 prompt —— 把 LLM 变成一把精准、不留情面的「心理手术刀」，对用户的自述做结构化分析：隐藏叙事、核心驱动情绪、自动化思维规则、反复掉进的毒性循环、根本矛盾，并用荣格原型把洞察人格化成阴影 / 未来自我两段独白。不是 AI 心理治疗（不提供安慰），而是一套自审协议：帮人用自己平时不敢、不会、做不到的方式审视自己。 关键特性：(1) 认识论诚实 —— 它没有关于你的客观真相，只有你给的自述，所以它产出的是「可供你拿真实经验去检验的锋利假设」，而不是判决；(2) 对模式锋利，对人不残忍；(3) 内建安全底线 —— 检测到急性危机信号（自杀/自伤/被虐待）会立刻放下手术刀转向真实求助资源；(4) 三阶段固定结构 + 字数上限 + 帕累托 80/20，输出稳定不发散。 适用：用户想被狠狠点醒、做深度自我剖析、看清自己反复犯的模式、走出内耗/焦虑/反复的困境；说「帮我分析一下我自己」「我是不是有什么问题」「为什么我总是…」「狠一点别安慰我」「给我做个心理手术」。 触发词：犀利博士、尖锐博士、尖锐医生、刀锋博士、dr sharp、心理手术刀、深度自审、狠狠分析我、别安慰我、人生教练 prompt、我总是、为什么我老是、戳穿我、把我看透、honest coach、psychological diagnosis、self-diagnosis。 不适用：用户处在急性心理危机中（→ 走安全底线，不做分析）；只想要情绪安慰和陪伴（这个 skill 故意不提供）；临床诊断 / 替代专业心理治疗（它不是，也永远不声称是）。

「三叉戟 / 三重心智」深度阅读 prompt —— 让 LLM 不对一篇文章 / 一份材料做泛泛概括，而是同时占据三个不重叠的视角陈述观点：建构者（抽出核心框架并往外推）、挑战者（最强反驳、最薄弱的一环在哪）、实践者（这对「我」的行动有什么具体启发）。逼模型同时当正方、反方、和顾问，产出比「summarize this」深得多的多视角分析。 关键特性：(1) 三个视角各有明确的产出契约 + 字数预算 —— 这是原版最缺的，避免发散成空话；(2) 每个视角开头一句话点明「我现在是哪个视角、它在优化什么」，把「让模型审视自己的思考」落成具体动作，而非玄学；(3) 单条消息直接喂材料，不要「你准备好了吗」的握手往返；(4) 模型无关，可换 DeepSeek/ChatGPT/Gemini；用途可从「分析文档」泛化到「评估一个决策」「学一个概念」「想清一个复杂问题」。 适用：用户给一篇长文 / 播客实录 / 论文 / 演讲 / 一份方案，想要的不是摘要，而是「帮我读透它」「多个角度看看」「这篇靠不靠谱」「这对我有什么启发」「批判性地分析这篇」「评估这个决策」。 触发词：三重心智、三叉戟、trident、深度阅读、读透这篇、多视角分析、批判性阅读、正反两方、这篇靠不靠谱、给我点启发、帮我想清楚、评估这个决策、challenger builder practitioner、deep reading、steelman。 不适用：用户真的只想要一句话摘要 / TL;DR（别上三视角）；需要逐句翻译或纯信息抽取（用对应工具）；材料太短不值得三个视角（一两段话直接聊）。

Use when a jewelry / 珠宝 / 文玩 / 翡翠 / 钻石 / 黄金 / 珍珠 e-commerce merchant needs marketing material from one product photo — produces a full XHS-ready bundle (12 marketing images + 6 copywriting styles + product analysis JSON) for finished pieces, or a design suite (8 images: hand-drawn sketch + 3D render + lifestyle + CAD sheet + ...) when the input is a raw gemstone / 原石 / 散石. Auto-detects finished-product vs raw-material. Trigger phrases: 珠宝营销素材, 生成珠宝商品图, 珠宝小红书素材, 珠宝海报, 上手图, 一饰多搭, 钻石/翡翠/珍珠/玉/古法金/月光石 marketing bundle, jewelry design from raw stone, 原石设计图, 散石变成品.

在 Claude Code 里主动引导你学任何东西的学习教练 —— 一门语言、一个技术主题、一项技能。 你想学 X,它带着你一步步学:轻校准你的水平 → 喂刚好略低于你水平的可理解输入(i-1)→ 让你用自己的话猜/复述(认出来即可,不要求背)→ 把你见过的词/概念在后续故意复现(撞见自然记)→ 正向小结"今天懂了" → 主动给下一小步。全程零评判、零考试、零背诵、零打卡 —— 第一目标是不痛苦、有小赢, 其次才是进度。领域无关内核 + 可选 per-domain playbook(英语是首个)。agent 自身就是导师,不调外部模型/API。 MUST trigger when the user says: "我想学X"、"教我X"、"带我学"、"陪我学"、"learning-coach"、"learning coach"、 "我对X一窍不通想入门"、"我一直学不好X"、"怎么开始学X"、"用 AI 教我X"。X 可以是英语、某个技术、乐器、任何技能/主题。

Set up a self-hosted sing-box proxy/VPN service end-to-end (VLESS-Reality / Hysteria2 / TUIC / VMess-WS / AnyTLS) on a fresh VPS — multi-protocol server, TLS cert via Cloudflare DNS-01, nginx subscription server, traffic stats. The node can be brought up two ways: a full fresh install, or by cloning an existing node. Use when the user wants to "set up a VPN on this VPS", "set up sing-box", "搭个梯子 / 搭一套翻墙服务", "add a proxy node", or "clone my proxy to another host".

从零陪伴非职业创作者完成游戏剧本的「全流程」skill —— 把模糊的灵感一步步引导成世界观、核心理念（主题）、故事主线框架、人物核心特质，再辅助写出具体的主线与支线剧情、场景级脚本。通用适配所有游戏类型（RPG/开放世界/视觉小说/Roguelike/模拟/解谜/竞技/沙盒…），其中【二次元手游抽卡游戏】覆盖最深。 核心特性：(1) 先校准创作者的真实水平（零基础 / 有想法没技巧 / 有直觉没训练 / 已能成稿）并据此调整辅导力度；(2) 中文为主、多语言支持；(3) 产出导向 —— 每个阶段都交付一份可用的实体文档（世界观画布/主题句/主线框架/人物档案/剧本），最终汇成一份"创作圣经"。 适用：用户想做一个游戏的剧情/剧本但不知从何下手；有个模糊点子（一个画面/一个角色/一种氛围/一个机制）想发展成完整故事；要写二游主线/角色PV/活动剧情/羁绊小剧情；要把已有世界观落地成具体场景对白脚本。 触发词：游戏剧本、剧情创作、写剧本、做个游戏故事、世界观怎么搭、主线框架、人物设定、人设、二游剧情、抽卡游戏剧情、角色剧情、主线剧情、活动剧情、邀约/羁绊剧情、卡池角色、版本剧情、PV文案、长线运营/版本规划、怎么写得催泪/好笑/反转、伏笔怎么埋、角色塌房、game script、game narrative、game story、worldbuilding、写不出来、不会写剧情、帮我把这个点子做成故事。 不适用：纯玩法/数值/关卡设计（非叙事）；已是职业编剧只要润色一段成稿（直接润色即可，无需全流程）；写小说/影视剧本而非游戏（可借用但本skill按游戏叙事的交互特性组织）。

横纵分析法（Horizontal-Vertical Analysis）深度研究 skill — 叙事型来龙去脉调研。由数字生命卡兹克提出，融合了索绪尔的历时-共时分析、社会科学的纵向-横截面研究设计、商学院案例研究法与竞争战略分析的核心思想。 当用户想要系统性研究一个产品、公司、概念、技术或人物时使用。核心是双轴分析：纵轴追踪从诞生到当下的完整生命历程（以叙事故事呈现），横轴在当下时间截面上与竞品/同类进行系统性横向对比，最后交叉两条轴产出独到洞察。最终产出一份排版精美的 PDF 研究报告。 适用场景：想从零搞懂一个东西的全貌（"帮我研究一下 X"、"X 是什么来头"、"X 的来龙去脉"、"帮我搞懂 X"），需要叙事型万字研究报告 + PDF 产出。 不适用：需要做决策判断的调研（"该选 A 还是 B"、"X 靠不靠谱" → 用 deep-research）、简单名词解释（"XX 是什么"）。 触发词：横纵分析、narrative research、研究一下、帮我搞懂、来龙去脉、什么来头、帮我了解、做个研究、全面了解、帮我摸清楚、horizontal-vertical analysis。

Use this skill when the user asks to "audit and fix", "review and fix all issues", "keep fixing until clean", "iterative fix cycle", or wants a thorough multi-round code quality pass. Runs code-reviewer audits and parallel fix agents in a loop until no high-confidence issues remain. Each round includes unit test + integration/e2e test authoring for fixed code and incremental doc updates. Also covers docs update, adhoc/orphan script cleanup, and proper file archiving.

Use this skill when the user asks to "do a full codebase audit", "sweep the whole codebase", "audit everything", "full health check", "clean up the whole project", "write architecture docs", "document the codebase", or wants a comprehensive one-time review of the entire project — not just recent changes. Covers full code audit, iterative fix loop with test authoring (unit + integration/e2e) and incremental docs, docs cleanup and reorganization, architecture documentation, and script archiving.

More Than Coding — full product workflow from concept to code. Iterates design/story first, generates aligned data, builds the app, then polishes. For games, demos, prototypes, and interactive experiences. Triggers: /mtc, build a game, make a demo, interactive experience.

Design completeness review for implementation plans. Rates 7 dimensions 0-10, detects AI slop patterns, builds interaction state tables. Triggers: design review, plan design review, check design completeness.

Default PDF generator. Render markdown, notes, or reports as professionally formatted PDFs with SF typography and clean layout. Use this whenever the user asks to "generate a PDF", "create a PDF", "make a PDF", or "convert to PDF" from any text or markdown source.

晚点 LatePost 风格 ZH 深度长文 skill——冷静判断 + 数据先行 + 判断式 header + 段尾落点 + 类比有锚 + 不端一面之词。 适用场景：行业分析、公司/产品拆解、趋势观察、一线观察专栏、技术深度访谈、财务战略分析、新政策/新事件解读。需要理性、深度、数据支撑、平衡视角的 deep dive。 不适用：纯个人体验/活人感叙事（用 khazix-writer）、企业官号 PR 文案（用 wechat-article-writer 或单独写）、教程/方法论（用 khazix-writer）。 字数无下限——短锐评/断言式 take（500-2000 字）也调本 skill，但走"轻量模式"（只跑 L1 硬性规则 + L4 voice 终审，跳过 L2 12 模式严格匹配）。短文也需要 AI slop 防御。 触发词：晚点风格、深度报道、行业拆解、公司分析、趋势观察、deep dive、写一篇深度长文、行业判断、写个 industry analysis、复盘 X、Y 公司怎么了。