all publishers

m1nga

@m1nga source repo

49 published skills

  1. ▌
    Podcast Revival · m1nga bundle
    Make a directed podcast video from supplied video, text, or mixed materials, with a requested duration, tone and point of view. Use for a new conversation, continuation, or source-based adaptation; a transcript summary alone does not require this skill.
    1 repo stars
  2. ▌
    Redbot · m1nga bundle
    Turn a selected saved post or RedBot library into a short answer, a checked next step or a usable coding prompt. Use for choosing a save, summarizing relevant source and comments, checking free-tool claims, or applying a saved method. Not a general web search or an automatic Xiaohongshu scraper.
    1 repo stars
  3. ▌
    Major Tom · m1nga bundle
    Turn a human directive into faithful, traceable execution by existing owners: separate intent, rules, reference documents and learned experience; select relevant skills and evidence; dispatch scoped work; verify outcomes; and improve future interpretation. Use for command-to-execution coordination, GrandControl/Major Tom workflows, or correcting repeated execution drift across existing agents. Ordinary single-step tasks need no orchestration ceremony.
    1 repo stars
  4. ▌
    Daily Brief · m1nga bundle
    Produce a source-backed US stock portfolio decision brief from holdings, new filings and current prices; evaluate hold, buy, trim, switch or wait, and investigate stock tips. Use for stock daily briefs, portfolio reviews and thesis checks, or when the user says 美股日报 / 看看我的持仓 / 这只该不该卖 / 财报出了帮我看 / 有人推荐了一只股票. Not general morning news, not crypto, not price alerts. Never executes trades.
    1 repo stars
  5. ▌
    Work Memory · m1nga bundle
    Keep useful work knowledge across conversations through selective, sourced records, project-scoped recall, outcome-tested lessons, and portable handover bundles. Use during ongoing project work, meaningful corrections or milestones, retrieval of prior decisions, and handover preparation; not for logging every chat or treating past choices as permanent rules.
    1 repo stars
  6. ▌
    Taskdock 2 · m1nga bundle
    TaskDock keeps a multi-session task in one portable folder (TASK.json plus README, STATE and PLAN) so a later session, another agent (Claude Code, Codex, ChatGPT) or another machine can resume it, find it after a move, and reorganize its files with a reversible plan. Use when work will outlive this conversation or hand off to someone else, or when the user asks to create, resume, find, hand off, or reorganize a task folder (任务文件夹 / 续接任务 / 接手 / 找回任务 / 整理任务文件). Not for one-off questions, tiny edits, or code that already lives in its own git repository.
    1 repo stars
  7. ▌
    Design Agent 2 · m1nga bundle
    Turn visual references and feedback into brand direction, web design, and verified implementation within the requested scope. Use for reference-led design collaboration, refining AI-generated interfaces, or when the user says 照这个参考做 / 品牌方向 / 改这个界面 / 网页设计 / 视觉稿 / 这版看着不对. Keep project assets and decisions in the project, not in this skill. Not for unrelated coding, marketing copy alone, or a first-time-user experience report (product-experience-officer).
    1 repo stars
  8. ▌
    Map Product System 3 · m1nga bundle
    Map a product's user journeys, capabilities, data, responsibilities, and missing paths at the scope requested: one feature, one flow, or the whole system. Use for system structure or architecture changes, before adding a feature that touches several parts, or when the user says 画一下产品结构 / 系统地图 / 用户旅程 / 这个功能缺什么路径 / 谁负责哪块 / 架构怎么改. Output names journeys, capabilities, owners, gaps, and buildable slices. Not for experiencing a product as a new user (product-experience-officer), not for organizing task files (taskdock), not for a repair-or-rebuild verdict (diagnose-project-rebuild), not for runtime automation loops.
    1 repo stars
  9. ▌
    Prompt Distill 2 · m1nga bundle
    Turn rough or dictated input into a reusable prompt when explicitly asked: 把这段整理成 prompt / 提炼成提示词 / 帮我写成可复用的指令 / turn this into a prompt. Preserve corrections, intent, examples, and constraints; keep the user's language. Do not intercept a task the user wants executed, and do not fire on ordinary requests that merely contain instructions.
    1 repo stars
  10. ▌
    Scenario Probe 2 · m1nga bundle
    Wind-tunnel any instruction text that configures AI behavior — a SKILL.md, system prompt, CLAUDE.md rule, agent definition, or plugin command — by projecting it into persona × scenario simulations before it ships. Derives personas from the owner's real contexts, predicts trigger decisions from the target harness’s pre-load routing metadata, walks the body line-by-line per scenario, and reports false-fires, missed triggers, sibling-skill collisions, stale-world failures, silent-failure paths, and stranger-usability gaps, each with line-anchored fixes. Use for "probe / wind-tunnel / stress-test this skill, prompt, or instruction", "will this description misfire?", 风洞 / 场景推演 / 会不会误触, or after writing/editing any SKILL.md or long-lived prompt. NOT for experiencing a runnable product with a UI (use product-experience-officer), NOT for interrogating a plan (thinking-partner's interview mode), NOT for evaluating model outputs (write an eval grader instead), NOT for unbuilt ideas (thinking-partner).
    1 repo stars
  11. ▌
    Thinking Partner 2 · m1nga bundle
    Think through an ambiguous choice or tradeoff when the user wants judgment rather than output: which option, whether it is worth doing, what they are missing. Also runs a relentless one-question-at-a-time interview when the user asks to be grilled before building ("grill me on this plan", "poke holes before I build", 拷问我 / 开工前把我问透 / 先问清楚再动手) and ends it with a locked, scoped plan. Chinese triggers: 帮我想想 / 我在纠结 / 哪个更好 / 拿个主意 / 值不值得做. For a clear task, execute it; do not turn a request into an interview unless asked, and do not fire on delegation ("you decide", 你看着办). Not for UX audits, product definition questionnaires, or recapping a conversation.
    1 repo stars
  12. ▌
    Prompt Craft 2 · m1nga
    Turn casual ideas into AI-ready prompts. Combines brand context, marketing knowledge, and Mingtao-specific communication patterns to produce prompts better than you'd write yourself. Fast by default, deep-audit on request or auto-escalated for high-stakes.
    1 repo stars
  13. ▌
    Coffee Brewing 2 · m1nga bundle
    Mingtao 的私人咖啡冲煮助手。看咖啡照片(豆袋 / 豆子本身 / 磨好的粉 / 萃出的浓缩液)→ 判断烘焙度与处理法 → 给出精确的研磨/水温/比例/时间参数和饮品配方(意式 espresso、冰美式、热美式、拿铁奶卡、V60 手冲),并通过「时间+口感」反馈帮你一步步调豆(dial-in),诊断萃取过度/不足,把调定的豆子存进豆子记忆,越用越准。硬件写死:Mazzer Philos 磨豆机、Hario V60、La Marzocco Linea Micra。 Use this skill whenever Mingtao shares a photo of coffee beans/bags/grounds/espresso, asks how to brew / grind / dial in any coffee, wants a recipe for espresso/americano/latte/pour-over, reports a shot tasting sour/bitter/under- or over-extracted, or mentions 冰美式/热美式/浓缩/手冲/磨豆/研磨度/养豆/萃取/拉花/奶卡/V60. Trigger even if they don't say the word "skill" — any coffee brewing or dial-in question counts.
    1 repo stars
  14. ▌
    Conclude Rounds 2 · m1nga
    Conclude the last N rounds of the CURRENT conversation, in Chinese by default, AND hand back 5 evidence-based insights for working more efficiently with Claude Code. Read back over the recent exchanges — both the user's questions AND the assistant's answers, including code written, decisions made, and progress — then give a tight conclusion (what was asked, what was actually done and verified vs merely proposed, current state, open decisions) followed by 5 concrete Claude-Code-usage insights drawn from patterns in those rounds. N defaults to 4 and is read from what the user says (前4轮 / past 3 / 前面5个来回 / 前面两个问题). This is an IN-conversation recap, read-only — it never changes files, and it is NOT a cross-session daily digest. Run whenever the user says 总结前面N轮 / 总结一下前面 / 小结一下 / 结论一下 / 回顾前面几轮 / 前4轮 / 这几轮做了啥 / conclude the last N rounds / recap the last few rounds / summarize what we just did / catch me up on this thread — even without naming the skill. Works identically in Claude Code and Codex.
    1 repo stars
  15. ▌
    Desktop Package 2 · m1nga
    Package the current task/session's content into ONE new folder on Ming's macOS Desktop for human pickup (查收) — distilled Chinese documents plus copies of files produced, organized so a cold reader gets the task, conclusion, and open items without the conversation. Folder name = one emoji + short Chinese topic (Ming's desktop habit); fixed entry file 📖 先看这个.md inside. Two modes — end-of-task packaging (把前面说的内容整理到桌面), and task-start filing (这个任务单独开个桌面文件夹, outputs accumulate into it during the task). Trigger whenever Ming says 整理到桌面 / 桌面开个文件夹 / 放到桌面我要查收 / 打包到桌面 / 桌面上给我留一份 / put this in a desktop folder / make a desktop folder for this task — even without naming the skill, even when the ask is the last sentence after a long setup. NOT for AI-session handoff (conversation-package), NOT for closing an iteration (iteration-close), NOT a desktop cleaner (整理桌面 is a different job) — it only ever ADDS one new folder and never touches existing Desktop items. Works identically in Claude Code and Codex.
    1 repo stars
  16. ▌
    Experience Pack 2 · m1nga bundle
    Two-layer experience system for any project: keep a per-project EXPERIENCE.md ledger (dated entries: what happened, what we did, what it cost), then distill it through an experience-vs-decision boundary test into portable, noun-free lessons that other projects can safely reuse — experience inherits, decisions die. Ships a reference pack distilled from a real 5-rebuild contamination saga (source-lock bundles, planner/executor/ verifier role isolation, zero-context probes, quarantined banned-noun lists, bounded repair loops, multi-AI reference-not-instruction triage) plus five prevention advices. Use when the user says 经验包 / 记录经验 / 复盘 / 经验文件 / 沉淀经验 / experience pack / lessons learned / post-mortem / record this experience / distill experience / "what did we learn" / wants a project to learn from another project's history, or after any expensive incident worth recording. NOT for closing an iteration (use iteration-close) and NOT a decision register — decisions stay in the project's own truth files.
    1 repo stars
  17. ▌
    Iteration Close 2 · m1nga
    Close a product iteration the safe-but-ruthless way, in ANY repo — not tied to one product. Distill this iteration's decisions into the repo's own record, DELETE superseded baggage (git history is the archive — no _archive dirs, no .bak, no legacy- renames), verify two SEPARATE readiness channels (a confirmed cloud upload AND a tested zero-context takeover probe), rehearse the one-command new-machine bootstrap, then tag and seed the next iteration. Adapts to each repo's conventions and creates the minimum where one is missing. Run when an iteration or slice is done/approved and the user says any of — 收口 / 迭代收口 / 迭代收尾 / 收尾 / 收工 / 封版 / 发版 / 甩包袱 / 清旧账 / 蜕壳 / 这个迭代做完了 / 收尾上云 / close the iteration / wrap up this iteration / ship this iteration / finalize this iteration / done with this iteration. Also handles a STRUCTURAL overhaul as an in-repo iteration, never a new-repo restart. Product-agnostic; works in Claude Code and Codex.
    1 repo stars
  18. ▌
    Map Product System 2 · m1nga bundle
    Turn rough product ideas, conversations, documents, or an existing codebase into an end-to-end product system map covering user journeys, capability boundaries, platform and customer mandates, data lifecycles, agent/service/human responsibilities, architecture, failure paths, governance, and delivery slices. Use when the user asks to understand the whole product, find missing functions or branches, establish product foundations, decide what belongs in code versus mandates or configuration, design multi-agent responsibilities, audit product completeness, or convert a product concept into an executable architecture. Trigger on phrases such as “看一下产品结构”, “梳理产品架构”, “检查用户旅程和功能分支”, “review the product structure”, or “map the whole product”.
    1 repo stars
  19. ▌
    Loop System Architect 2 · m1nga bundle
    Design, audit, repair, and operationalize reliable agent loops with explicit goals, triggers, persistent state, orchestration, independent verification, failure recovery, open skill discovery or creation, experience distillation, controlled self-improvement, and retirement conditions. Use when the user asks to build a loop, autonomous or recurring agent workflow, multi-agent operating system, self-improving process, knowledge/experience distillation pipeline, zero-context recovery flow, permanent update/replacement loop, or to turn a prose plan or repeated prompt into a genuinely executable loop.
    1 repo stars
  20. ▌
    Diagnose Project Rebuild 2 · m1nga bundle
    Diagnose a confused, inherited, partially lost, repeatedly rebuilt, or direction-changing project and establish the safest evidence-based foundation for what comes next. Use when the user says a project is chaotic, polluted by old decisions, only an older version survives, the original purpose must be rediscovered, they are unsure whether to recover/repair/rebuild/continue, or they want to restart without importing the prior project's strategy. Adapt the route to the project's actual failure mode rather than forcing a universal sequence.
    1 repo stars
  21. ▌
    Product Experience Officer 2 · m1nga
    资深产品体验官 — experience a product-in-development as a zero-experience first-time user, then report to the PM with prioritized findings, fix recommendations, and follow-ups. Covers product experience, design, visual, interaction, copy, emotion — everything a cold stranger would feel. Two modes — analyze screenshots the user provides, or run/drive the product live (web via browser tools, CLI/TUI via terminal, native via computer-use). Use whenever Ming shares screenshots of a product he's building and wants reactions, or says anything like 体验一下 / 用户视角 / 从0经验的角度 / 体验官 / 帮我试试这个产品 / walkthrough / first-time user experience / UX audit / 看看新用户会怎么想 / 这个流程顺不顺 — even if he doesn't say the skill name. Trigger for any "experience my product and give feedback" request, screenshots included or not.
    1 repo stars
  22. ▌
    Think4ming · m1nga bundle
    Ming's automatic thinking-partner mode for ambiguous framing, open exploration, trade-off analysis, or decisions where misreading the need would materially change direction. Trigger semantically when the primary deliverable is better understanding, option-building, or judgment, even if Ming does not name the skill. Do not use it as a global wrapper for fact research, execution/building, artifact review or diagnosis, summarization, extraction, translation, monitoring, or status work with a clear target. For mixed requests, apply it only to the unresolved decision fork, then continue with the relevant research, execution, or review workflow in the same task. Never ask Ming to decide whether this skill should be invoked.
    1 repo stars
  23. ▌
    Prompt4ming · m1nga bundle
    Clarify Mingtao's intended outcome from rough, voice-dictated, mixed Chinese-English, or self-correcting input while preserving his meaning and direct style. Use when Ming explicitly asks to optimize, clean up, clarify, or turn his words into a reusable prompt, or when ambiguity in dictated input prevents reliable execution. Do not invoke merely because his request is conversational or long.
    1 repo stars
  24. ▌
    Mixtape · m1nga bundle
    Make the user an authored playlist and deliver a verified one-click Soundiiz import link for Apple Music, YouTube Music, Spotify, Deezer, or Tidal without platform API keys. Supports quick orders from a scene or mood, expansion from seed songs, and taste-profile curation from playlists the user shares. Use for "make me a playlist," "songs for late-night coding," 给我来个歌单, 根据这几首帮我扩展, 找类似的歌, 分析我的听歌口味, or any request for songs to listen to. Not for audio narration, transferring an existing playlist unchanged, or controlling playback.
    1 repo stars
  25. ▌
    Grilling · m1nga bundle
    Interview the user relentlessly about a plan or build, then produce a locked, scoped plan before any execution. Trigger when the user explicitly asks to be questioned or stress-tested — "grill me on this plan", "poke holes in this before I build it", "interview me until this is fully scoped", "ask me everything you need before writing code", "don't let me start until the plan is tight"; Chinese triggers — 拷问我 / 盘问这个方案 / 开工前把我问透 / 先问清楚再动手. Also trigger for an explicit build the user wants scoped through questioning before execution — only when they show willingness to answer questions. Do NOT trigger on delegation or authorization phrases ("just do it", "you decide", "你看着办") — those grant execution, not an interview. For lightweight feedback on a plan, or open-ended thinking with no concrete build to scope, yield to the thinking-partner skill. NOT for mock interviews or quiz practice ('grill me on React questions').
    1 repo stars
  26. ▌
    Taskdock · m1nga bundle
    TaskDock gives ongoing tasks a portable Desktop workspace: plan the work, organize task files, track decisions and next actions, and resume after a folder moves. Use for complex or continuing work, or when the user asks to create, resume, find, or reorganize a task folder.
    1 repo stars
  27. ▌
    Idea Probe · m1nga bundle
    Wind-tunnel a new idea before anything gets built. When the user brings a product, skill, site, or feature idea and asks for a pre-build test — "simulate some users and test this", "probe this idea", 帮我模拟用户测测 / 新想法:先别建,推演一下 — derive 4–7 personas from the idea's claimed audience with forced diversity slots, then run each through honest first-contact simulation: discovery → first-glance reading (misreads recorded verbatim) → start or not → first friction → return or not. Output: BUILD/NARROW/RETHINK verdict with assumption debt, problems ranked by breadth × fatality, a fix each, a revised one-pager, and a reality bridge: real-user tests with kill criteria for the deadliest findings — simulation generates hypotheses, never market evidence. NOT for a bare "这个想法怎么样 / what do you think" (answer directly, or thinking-partner), NOT for instruction text or SKILL.md files (scenario-probe's seat), NOT for a runnable product (product-experience-officer's seat), NOT for technical feasibility (research, not personas).
    1 repo stars
  28. ▌
    Product 5w · m1nga bundle
    Interrogate a product's definition with five questions, at any stage — idea, mid-build, or shipped — asking past behavior, never future opinion. WHO (by behavior not demographics; who it's NOT for; user vs buyer; first 10 users), WHAT job it does (and deliberately doesn't; do-nothing as first competitor), WHEN as context and trigger (struggling moment, frequency, tense-matched), HOW it gets used AND found (distribution is half of HOW; ultimate test: an offer), WHY it matters / why now / why you. Answers tagged verified/inferred/assumed; each W splits desk reasoning vs field proof; output: DEFINED/NARROW/REDEFINE verdict + validation debt, cheapest-next-verification list, named contradictions. Use for "run a 5W on this", "who is my product actually for", "survey my product before launch", 给我的产品做个调研 / 这个产品的 who what why 帮我过一遍 / 上线前做个基本盘调研 / 产品定义审计. NOT persona simulation (idea-probe), NOT architecture mapping (map-product-system), NOT hands-on audit (product-experience-officer), NOT technical feasibility.
    1 repo stars
  29. ▌
    Side Quest · m1nga bundle
    Park a mid-flow thought without breaking your main thread — capture it in one line, a background agent works it while you keep going, and results land in an inbox outside the chat, keeping main context clean both ways. Ledger-first (a parked thought can never be silently lost), draft-only (side agents never send, commit, or touch your working tree), receipts carry a one-line verdict. Fires without confirmation on explicit markers — "sq:", "side quest:", "background this", "park this", "not now but:", 支线 / 先别进主线 / 后台弄 — and on queue check-ins ("sq list", "sq retry", "sq kill", "what came back from my side quests?"). Inferred phrasings get a one-line confirm, defaulting to the main thread. NOT for blocking prerequisites the main task needs next, NOT for ten-second questions (answered inline), NOT for reminders (routed to a scheduler, never faked as work), NOT for merely setting a topic aside with no work attached (这个先放一边), NOT a way to send or do anything irreversible.
    1 repo stars
  30. ▌
    Design Agent · m1nga bundle
    Turn visual references and feedback into brand direction, web design, and verified implementation within the requested scope. Use for reference-led design collaboration or refining AI-generated interfaces; not for unrelated coding or marketing copy alone.
    1 repo stars
  31. ▌
    Extend First · m1nga bundle
    Check for reusable skills, agents, or automations before creating a new capability. Recommend extending, combining, or building; skip for ordinary tasks and explicit requests to bypass the check.
    1 repo stars
  32. ▌
    One Sentence · m1nga bundle
    Answer with ONE crafted sentence instead of a paragraph — for explaining concepts, defining terms, and writing self-intros, taglines, loglines, or positioning lines. Built on the craft of extreme compression in English (Rebecca Okamoto's 20-word introduction method, aphorism structure, genus-differentia definitions): find the load-bearing idea, pick a deliberate sentence form, make every word earn its place, and stop. Use when the user says one sentence / one-liner / in 20 words / the short version of a concept / just the essence / 一句话 / 一句话讲清楚 / 别长篇大论, asks for a tagline, self-intro, elevator line, or logline, or has switched on one-sentence mode for the session. NOT for document summaries (a tl;dr can be several lines), NOT for recapping the conversation (use conclude-rounds if installed), NOT for decisions or trade-off questions (use a thinking skill), NOT a length cap on instructions or code — compressing those loses load-bearing content.
    1 repo stars
  33. ▌
    Prompt Craft · m1nga bundle
    Turn rough notes and revisions into reusable marketing prompts grounded in product facts, audience and brand voice. Use when the user asks to build or improve a prompt for ads, posts, campaigns or other marketing content. Direct copywriting and general marketing questions remain ordinary work.
    1 repo stars
  34. ▌
    Coffee Brewing · m1nga bundle
    A dial-in coach for home coffee. Reads photos of beans, bags, grounds, or espresso shots to judge roast level and process; gives ONE concrete grind / water temp / ratio / time recommendation; recipes for espresso, iced/hot americano, latte/flat white, and V60 pour-over; converges shot by shot on time + taste feedback, diagnosing under- vs over-extraction. Remembers every dialed-in bean (in ~/.coffee/, outside the package) so a repeat bag gets its proven parameters instantly. Trigger when the user shares coffee photos, asks how to brew, grind, or dial in, wants a recipe, or reports a shot tasting sour, bitter, weak, or "off" — e.g. "my espresso tastes sour", "how to grind these beans?", "V60 recipe for a light roast" — or mentions 冰美式 / 热美式 / 浓缩 / 手冲 / 磨豆 / 研磨度 / 调豆 / 养豆 / 萃取 / 拉花 / 奶卡 / V60. Adapts to any grinder/machine via one-time hardware setup; covers moka pot and French press via roast baselines. Do NOT trigger on metaphorical coffee or extraction ("data extraction", "wake up and smell the coffee").
    1 repo stars
  35. ▌
    Prompt Distill · m1nga bundle
    Turn rough or dictated input into a reusable prompt when explicitly asked. Preserve corrections, intent, examples, and constraints. Do not intercept a task the user wants executed.
    1 repo stars
  36. ▌
    Scenario Probe · m1nga bundle
    Wind-tunnel any instruction text that configures AI behavior — a SKILL.md, system prompt, CLAUDE.md rule, agent definition, or plugin command — by projecting it into persona × scenario simulations before it ships. Derives personas from the owner's real contexts, predicts trigger decisions from the target harness’s pre-load routing metadata, walks the body line-by-line per scenario, and reports false-fires, missed triggers, sibling-skill collisions, stale-world failures, silent-failure paths, and stranger-usability gaps, each with line-anchored fixes. Use for "probe / wind-tunnel / stress-test this skill, prompt, or instruction", "will this description misfire?", 风洞 / 场景推演 / 会不会误触, or after writing/editing any SKILL.md or long-lived prompt. NOT for experiencing a runnable product with a UI (use product-experience-officer), NOT for interrogating a plan (use grilling), NOT for evaluating model outputs (use write-judge-prompt), NOT for unbuilt ideas (idea-probe).
    1 repo stars
  37. ▌
    Conclude Rounds · m1nga bundle
    Recap the last N rounds of the CURRENT conversation plus up to 5 evidence-based insights for working faster with your coding agent. Separates done-and-verified from done-but-unproven from merely-proposed — a proposal never reads as a completion. N defaults to 4. In-conversation, read-only — not a cross-session digest, never edits files. Trigger on "recap the last few rounds", "what did we just do?", "summarize what we just did", "catch me up on this thread", "what actually got done vs just discussed?", or Chinese 总结前面N轮 / 总结一下前面 / 结论一下 / 前4轮 / 小结一下 / 回顾前面几轮 / 这几轮做了啥 — even without naming the skill. Replies in the language of the user's recent messages.
    1 repo stars
  38. ▌
    Desktop Package · m1nga bundle
    Collect task results into one reviewable Desktop folder, at task start or completion. Copy artifacts and summarize decisions; do not reorganize the Desktop or replace the project’s working files.
    1 repo stars
  39. ▌
    Experience Pack · m1nga bundle
    Two-layer experience system for any project: keep a per-project EXPERIENCE.md ledger (dated entries: what happened, what we did, what it cost), then distill it through an experience-vs-decision boundary test into portable, noun-free lessons that other projects can safely reuse — experience inherits, decisions die. Ships a reference pack distilled from a real five-rebuild contamination saga plus five prevention advices. Use for project-level retrospectives and incident recording, or right after an incident that cost real time or money — "let's do a post-mortem on this project", "write down what we learned before we forget", "record this incident so we don't repeat it", "what lessons from the last project apply here", "start an experience log for this repo"; Chinese: 经验包 / 记录经验 / 项目复盘 / 沉淀经验 / 经验文件 / 这次的教训. NOT on reviewing the current conversation's rounds (use conclude-rounds), NOT on closing an iteration (use iteration-close), and NOT a decision register — decisions stay in the project's own truth files.
    1 repo stars
  40. ▌
    Iteration Close · m1nga bundle
    Close out a development iteration, sprint, or milestone in any repo, distill chat-born decisions into the repo's own record, then shed superseded files — git history is the archive (no _archive dirs, .bak, or legacy- renames), the full deletion list needs the user's explicit yes before anything is removed, and untracked files are quarantined, never rm'd. Verify two separate readiness channels (a confirmed upload AND a tested zero-context takeover probe), rehearse the one-command bootstrap on a fresh clone, then tag and seed the next iteration. Runs the cleanup phase alone when the user only wants dead files cleared. Trigger on: "close out this iteration", "wrap up this sprint and clean out the dead files", "could a fresh machine take this over?", "clear the superseded files from the repo", "milestone's done — tidy and tag it", 收口 / 迭代收口 / 迭代收尾 / 收尾 / 封版 / 甩包袱 / 清旧账 / 蜕壳 / 这个迭代做完了 / 可以发版了 / 这版可以发版了 / 收尾上云. NOT a casual end-of-day sign-off; NOT how-to-publish-a-package questions. Works in Claude Code and Codex.
    1 repo stars
  41. ▌
    Voice Extractor · m1nga bundle
    Extract a user's real writing voice from 5-20 samples into a measurable fingerprint (voice.yaml), then check and enforce drafts against recorded habits and explicit style rules. Measures voice with named stylometry lenses (Burrows's Delta function words, MATTR lexical diversity, sentence-length burstiness, Biber register, opener profile, punctuation rates) and gates drafts as numeric bands, not vibes. Use when the user wants writing to sound like themselves or flags AI-sounding output: "make it sound like me", "my drafts sound AI-written", "capture my writing voice", "this doesn't sound like something I'd write", "set up / refresh / check my voice fingerprint", "像不像我", "(文案/草稿)太 AI 腔了", "这不像我写的". Also use when a co-installed drafting skill needs sender-voice constraints before returning copy. Do not use for generic proofreading, grammar fixes, or tone rewrites unrelated to a personal voice. Not a humanizer or bot-detector evasion tool.
    1 repo stars
  42. ▌
    Thinking Partner · m1nga bundle
    Think through ambiguous choices and tradeoffs when the user wants judgment or exploration. For a clear task, execute it; use this only for a decision that could change direction.
    1 repo stars
  43. ▌
    Map Product System · m1nga bundle
    Map a product’s user journeys, capabilities, data, responsibilities, and missing paths. Use for system structure or architecture changes; scale the map to the requested scope.
    1 repo stars
  44. ▌
    Validate Evaluator · m1nga bundle
    Calibrate an LLM judge against human labels using train/dev/test splits, TPR/TNR, and bias correction (Rogan-Gladen), so its scores can be trusted. Use after writing a judge prompt (companion skill write-judge-prompt, if installed) and before believing any number the judge produces. Triggers on requests like "is my judge accurate", "can I trust these scores", "validate my evaluator", "my LLM judge keeps disagreeing with me", "measure judge alignment with human labels". 中文触发:裁判准不准 / 这些分数可信吗 / 校准评估器 / 验证 LLM 评审. Do NOT use for code-based evaluators — those are deterministic; test them with standard unit tests. Requires human-labeled traces (or raw traces to label — a minimal labeling walkthrough is included) and a Python environment (numpy, scikit-learn; optionally judgy).
    1 repo stars
  45. ▌
    Write Judge Prompt · m1nga bundle
    Design a binary Pass/Fail LLM-as-Judge evaluator for one specific subjective failure mode that code-based checks cannot catch (tone, faithfulness, relevance, completeness). Use when the user says things like "write a judge prompt", "I need an LLM to grade my chatbot's outputs", "how do I automatically evaluate tone / faithfulness / relevance", "build an evaluator for my agent's responses", or "turn my error analysis into an eval". 中文触发:写裁判提示词 / LLM 评审 / 自动评估主观质量 / 帮我做个评估器. Do NOT use when the failure mode can be checked with code (regex, schema validation, execution tests) — write that check instead. Do NOT use for one-off manual review of a single output; this skill builds a reusable evaluator. For validating or calibrating a finished judge, use its companion skill validate-evaluator, if installed; for enforcing a personal voice fingerprint ("does this sound like me"), use voice-extractor, if installed.
    1 repo stars
  46. ▌
    Loop System Architect · m1nga bundle
    Design, audit, repair, and operationalize reliable agent loops with explicit goals, triggers, persistent state, orchestration, independent verification, failure recovery, open skill discovery or creation, controlled self-improvement, and retirement conditions. Use to build a loop, an autonomous or recurring agent workflow, a multi-agent operating system, a self-improving process, a zero-context recovery flow, a permanent update/replacement loop, or to turn a prose plan or repeated prompt into a genuinely executable loop. Trigger on natural phrasings such as "I keep running this prompt by hand — automate it", "make this agent run on its own and check its own work", "why does my automation silently fail", and Chinese phrasings such as "把这个流程变成自动循环", "帮我设计一个自动跑的 agent 工作流". For a plain scheduled reminder or interval re-run without state or verification, use the host scheduler (cron/schedule tools) instead. For static product architecture and responsibility contracts, use map-product-system.
    1 repo stars
  47. ▌
    Publish Skill Product · m1nga bundle
    Finish self-authored skills through user-perspective review, comparable-skill research, tested improvements, and verified GitHub publication. Use when creating, materially updating, or releasing an owned skill, and for skill discovery experiments.
    1 repo stars
  48. ▌
    Diagnose Project Rebuild · m1nga bundle
    Diagnose a confused, inherited, partially lost, repeatedly rebuilt, or direction-changing project and establish the safest evidence-based foundation for what comes next. Use when the user says things like "I inherited a mess", "should I just rewrite this from scratch", "this project is chaotic / polluted by old decisions", "can this still be saved", "only an older version survives", "I keep rebuilding this and the old plan keeps coming back", "fix it or start over?", or wants to restart without importing the prior direction's strategy; Chinese triggers: 乱了 / 被污染 / 推倒重来 / 还能不能救 / 接手的烂摊子 / 修还是重建. Adapts the route (recover / clarify / repair / rebuild / continue / hold) to the project's actual failure mode rather than forcing a universal sequence. NOT for tidying files or folders, and NOT for mapping a healthy product's structure (use map-product-system).
    1 repo stars
  49. ▌
    Product Experience Officer · m1nga bundle
    Experience a product-in-development as a zero-context first-time user, then report to the person who built it with a verdict, prioritized findings, concrete fix recommendations, and follow-ups. Covers comprehension, onboarding, core loop, interaction, visual design, copy, and emotion — everything a cold stranger would feel. Two modes — analyze screenshots the user provides, or run the product live (web, CLI/TUI, or native). Trigger on requests like "walk through my app as a first-time user", "does this onboarding make sense?", "UX audit these screenshots", "pretend you've never seen this and try it", "would a new user understand this?", or Chinese phrasings 体验一下 / 用户视角 / 从0经验的角度 / 体验官 / 帮我试试这个产品 / 看看新用户会怎么想 / 这个流程顺不顺(产品体验语境). Any "experience my product and give me feedback" request counts, screenshots included or not. NOT for debugging a specific error shown in a screenshot, code review, or fixing bugs — this skill evaluates the experience, it does not repair the build.
    1 repo stars