[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7DJXP7AJRBF4R6PNJCJDMR3":3,"skill-same:01M7DJXP7AJRBF4R6PNJCJDMR3":55},{"skillNo":4,"name":5,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":21,"frontMatter":22,"refName":23,"commitSha":24,"skillDirectory":25,"forksCount":13,"repositoryHtmlUrl":26,"repositoryHomepageUrl":6,"ownerHtmlUrl":27,"ownerBlogUrl":28,"ownerDescription":29,"updatedAt":30,"install":31,"resources":44},"01M7DJXP7AJRBF4R6PNJCJDMR3","write-human",null,"marliechorgan\u002Fclaude-setup","marliechorgan","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F101947869?v=4","01M7DJTFHCBEX24MHN2QNMNVN3","Charlie Morgan","01M7DJTFJ3G50SSPPEFTHXJ7Q3",0,"MIT","https:\u002F\u002Fgithub.com\u002Fmarliechorgan\u002Fclaude-setup\u002Fblob\u002Fmain\u002Fplugins\u002Fskills\u002Fskills\u002Fwrite-human\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: write-human\ndescription: Draft or edit writing a person will send or publish as themselves: texts, WhatsApp and DMs, emails, replies, LinkedIn messages, introductions and first approaches, client updates, cover notes, posts and short professional documents, so it sounds like them, not like an AI. Use whenever the user asks to write, reply to, rewrite, polish, shorten or \"make this sound less like AI\", or asks \"what should I say to X\". Drafts from the sender's own past messages, leads with the reader, cuts preview lines, stock phrases and research the reader never shared, and checks \"tomorrow\" and every other time word against the day it is sent. Drafts only; the user sends.\nlicense: MIT\n---\n\n# Write human\n\nWrite something the user can send as themselves. Natural writing comes from having something specific to say to a specific reader, not from vocabulary bans or deliberate roughness. The user's own wording and edits outrank every default here.\n\n## Start with the reader\n\nWork out who will read it, what they already know, what it must achieve, and where they will read it (a phone notification, an inbox, a printed page). Ask only when a missing fact would change the draft. Pick the form the job needs: a warm reply, a first approach and a technical update are different things.\n\nIf the user supplies a draft, keep their meaning, emphasis and good phrasing, and edit at the level they asked for.\n\nRead the live thread before replying when its state matters: a local draft or a sent-items log is not proof of what was actually sent or received.\n\n## Draft in their voice, from their words\n\nThe user's voice is in their sent messages, not in a general idea of \"friendly\" or \"professional\". For anything personal, ask for (or find, if you have access) three to five of their recent messages **to this person** and draft from those: length, capitals, punctuation, emoji, how they sign off. People write very differently to a grandparent, a colleague and a sibling. If no samples are available, write plainly and briefly, and say which parts are guesses.\n\nWhen the user adds a line of their own, keep it exactly as written.\n\n## What makes it read as a person\n\n- **Lead with them, not with you.** Open on their news, their question or what they said last time. Your result is often the least interesting part to the reader.\n- **Answer the last unanswered thing first.** If they made a joke, pick it up.\n- **Say the real thing plainly.** If someone is going through something, ask about it directly and keep it light; don't circle it or get earnest.\n- **Hand them the choice.** Offer a loose window and let them pick the time and place.\n- **One line about them** where it fits: their news, their work, a bit of praise that is true.\n- **Leave out what you found by digging.** A detail from their profile, an event page or another chat reads as surveillance. Use only what they told the sender.\n- **Warmth grounded in something real.** Plain stock warmth (\"thanks for your patience\") is fine when it fits an actual event. Don't invent gratitude, excitement or a favour to thank them for, and never supply a reason, news or an apology the sender did not give.\n- **Keep dates short, and check them on the day the draft goes.** In a message that goes the day it is written, a plan for the next day can say \"tomorrow\". A plan in a draft that may sit gets a short named day, \"Friday 13th\", not \"tomorrow\" or \"this weekend\"; add the month only when the day alone could mean two dates. Something already past stays loose, \"at the end of last week\", not a full weekday and date. Drafts often sit for a day or two, so re-read every time word, the loose ones too, against the day it is actually sent.\n\n## What to cut\n\n- A first line that previews the message (\"A quick update and one bigger question\").\n- A second sentence that shrinks a gracious out (\"No problem if not. It works fine without it.\").\n- Limitations the reader doesn't need in order to say yes.\n- Flourishes that read as boasts, over-offering (\"happy to do a panel, a talk, or whatever suits\"), and a PS bolted onto a first approach. One ask per message.\n- Presuming the reader's state of mind (\"as I'm sure you've noticed\"), and contrarian jabs.\n- Something the reader already has from an earlier message or from the product itself.\n\nLength is not the goal. Real edits lengthen as often as they shorten; what goes is logistics, research and devices, and what comes in is something about the person. Revise by cutting: a second pass deletes, it doesn't polish.\n\n## By kind of writing\n\n- **Texts and DMs:** usually two to five short messages rather than one paragraph. Show each separate message as its own blockquote so the user can paste them one by one. End on the practical line.\n- **First approaches and introductions:** describe a company by what it does, not where it is; for a non-technical or senior reader, purpose and impact beat technique, and one checkable credential beats a product description. Lead with what people can picture using, not with the tool the writer is proudest of.\n- **Emails and updates in any channel:** name the one point the message has to land, open with a short paragraph that says it head-on, then give one or two proof points: concrete facts that make it believable. The ask or next step comes after and is easy to find. One topic per email, and a second email names the first; a disclosure or apology in one plain sentence, not a confession. Introduce anything embedded (a template, a quoted reply, a table) in one sentence that says what it is and where it came from. When the reader must act on a decision, give an instruction they can relay (\"Use version 2 for this batch; flag uncertain words\") and say plainly if it is still pending.\n- **Technical and research updates:** real system names, exact numbers with their units and comparison (\"error rate fell from 30% to 16%\", not only \"halved\"), and what remains uncertain. Respect what the reader already knows; don't explain their own project back to them.\n- **Posts and shared documents:** follow the conventions already on the page or in the thread.\n\n## Check, then hand over\n\nIn Claude Code or anywhere Python runs, check the drafts before showing them:\n\n```bash\npython3 scripts\u002Fdraft_lint.py --markdown reply.md       # every blockquote is one message\npython3 scripts\u002Fdraft_lint.py --channel email draft.txt # one email\n```\n\nIt flags em dashes, stock phrases, \"not X but Y\" constructions, header-style labels, relative dates and over-long texts. (The phrase lists live in the script on purpose: naming them here would prime them.) If a flag is wrong for this reader because the sender genuinely writes that way, keep it and say so.\n\nShow the drafts ready to paste, with no commentary on the writing process. This skill drafts; the user sends. When they edit a draft and ask for another pass, carry their edits forward rather than regenerating from scratch.\n","# Write human\n\nWrite something the user can send as themselves. Natural writing comes from having something specific to say to a specific reader, not from vocabulary bans or deliberate roughness. The user's own wording and edits outrank every default here.\n\n## Start with the reader\n\nWork out who will read it, what they already know, what it must achieve, and where they will read it (a phone notification, an inbox, a printed page). Ask only when a missing fact would change the draft. Pick the form the job needs: a warm reply, a first approach and a technical update are different things.\n\nIf the user supplies a draft, keep their meaning, emphasis and good phrasing, and edit at the level they asked for.\n\nRead the live thread before replying when its state matters: a local draft or a sent-items log is not proof of what was actually sent or received.\n\n## Draft in their voice, from their words\n\nThe user's voice is in their sent messages, not in a general idea of \"friendly\" or \"professional\". For anything personal, ask for (or find, if you have access) three to five of their recent messages **to this person** and draft from those: length, capitals, punctuation, emoji, how they sign off. People write very differently to a grandparent, a colleague and a sibling. If no samples are available, write plainly and briefly, and say which parts are guesses.\n\nWhen the user adds a line of their own, keep it exactly as written.\n\n## What makes it read as a person\n\n- **Lead with them, not with you.** Open on their news, their question or what they said last time. Your result is often the least interesting part to the reader.\n- **Answer the last unanswered thing first.** If they made a joke, pick it up.\n- **Say the real thing plainly.** If someone is going through something, ask about it directly and keep it light; don't circle it or get earnest.\n- **Hand them the choice.** Offer a loose window and let them pick the time and place.\n- **One line about them** where it fits: their news, their work, a bit of praise that is true.\n- **Leave out what you found by digging.** A detail from their profile, an event page or another chat reads as surveillance. Use only what they told the sender.\n- **Warmth grounded in something real.** Plain stock warmth (\"thanks for your patience\") is fine when it fits an actual event. Don't invent gratitude, excitement or a favour to thank them for, and never supply a reason, news or an apology the sender did not give.\n- **Keep dates short, and check them on the day the draft goes.** In a message that goes the day it is written, a plan for the next day can say \"tomorrow\". A plan in a draft that may sit gets a short named day, \"Friday 13th\", not \"tomorrow\" or \"this weekend\"; add the month only when the day alone could mean two dates. Something already past stays loose, \"at the end of last week\", not a full weekday and date. Drafts often sit for a day or two, so re-read every time word, the loose ones too, against the day it is actually sent.\n\n## What to cut\n\n- A first line that previews the message (\"A quick update and one bigger question\").\n- A second sentence that shrinks a gracious out (\"No problem if not. It works fine without it.\").\n- Limitations the reader doesn't need in order to say yes.\n- Flourishes that read as boasts, over-offering (\"happy to do a panel, a talk, or whatever suits\"), and a PS bolted onto a first approach. One ask per message.\n- Presuming the reader's state of mind (\"as I'm sure you've noticed\"), and contrarian jabs.\n- Something the reader already has from an earlier message or from the product itself.\n\nLength is not the goal. Real edits lengthen as often as they shorten; what goes is logistics, research and devices, and what comes in is something about the person. Revise by cutting: a second pass deletes, it doesn't polish.\n\n## By kind of writing\n\n- **Texts and DMs:** usually two to five short messages rather than one paragraph. Show each separate message as its own blockquote so the user can paste them one by one. End on the practical line.\n- **First approaches and introductions:** describe a company by what it does, not where it is; for a non-technical or senior reader, purpose and impact beat technique, and one checkable credential beats a product description. Lead with what people can picture using, not with the tool the writer is proudest of.\n- **Emails and updates in any channel:** name the one point the message has to land, open with a short paragraph that says it head-on, then give one or two proof points: concrete facts that make it believable. The ask or next step comes after and is easy to find. One topic per email, and a second email names the first; a disclosure or apology in one plain sentence, not a confession. Introduce anything embedded (a template, a quoted reply, a table) in one sentence that says what it is and where it came from. When the reader must act on a decision, give an instruction they can relay (\"Use version 2 for this batch; flag uncertain words\") and say plainly if it is still pending.\n- **Technical and research updates:** real system names, exact numbers with their units and comparison (\"error rate fell from 30% to 16%\", not only \"halved\"), and what remains uncertain. Respect what the reader already knows; don't explain their own project back to them.\n- **Posts and shared documents:** follow the conventions already on the page or in the thread.\n\n## Check, then hand over\n\nIn Claude Code or anywhere Python runs, check the drafts before showing them:\n\n```bash\npython3 scripts\u002Fdraft_lint.py --markdown reply.md       # every blockquote is one message\npython3 scripts\u002Fdraft_lint.py --channel email draft.txt # one email\n```\n\nIt flags em dashes, stock phrases, \"not X but Y\" constructions, header-style labels, relative dates and over-long texts. (The phrase lists live in the script on purpose: naming them here would prime them.) If a flag is wrong for this reader because the sender genuinely writes that way, keep it and say so.\n\nShow the drafts ready to paste, with no commentary on the writing process. This skill drafts; the user sends. When they edit a draft and ask for another pass, carry their edits forward rather than regenerating from scratch.\n",{},"main","5d92f092e7cdba2bb603c1761587541dded44648","plugins\u002Fskills\u002Fskills\u002Fwrite-human","https:\u002F\u002Fgithub.com\u002Fmarliechorgan\u002Fclaude-setup","https:\u002F\u002Fgithub.com\u002Fmarliechorgan","https:\u002F\u002Fjustaniceguy.ai","AI deployment engineer, London. Evals and safety guards for coding agents; fixes merged in Inspect and ControlArena. Voice agents and speech models.","2026-11-14 09:16:31",{"source":7,"skillName":5,"commands":32,"prompt":42,"zipAvailable":43},[33,36,39],{"tool":34,"command":35},"npx","npx skills add marliechorgan\u002Fclaude-setup --skill write-human",{"tool":37,"command":38},"bunx","bunx skills add marliechorgan\u002Fclaude-setup --skill write-human",{"tool":40,"command":41},"pnpm","pnpm dlx skills add marliechorgan\u002Fclaude-setup --skill write-human","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fmarliechorgan\u002Fclaude-setup\u002Fblob\u002Fmain\u002Fplugins\u002Fskills\u002Fskills\u002Fwrite-human\u002FSKILL.md",true,[45,51],{"path":46,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":49,"sourceHtmlUrl":50},"scripts\u002Fdraft_lint.py","SCRIPT","脚本",7918,"https:\u002F\u002Fgithub.com\u002Fmarliechorgan\u002Fclaude-setup\u002Fblob\u002Fmain\u002Fplugins\u002Fskills\u002Fskills\u002Fwrite-human\u002Fscripts\u002Fdraft_lint.py",{"path":52,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":53,"sourceHtmlUrl":54},"scripts\u002Ftest_draft_lint.py",2700,"https:\u002F\u002Fgithub.com\u002Fmarliechorgan\u002Fclaude-setup\u002Fblob\u002Fmain\u002Fplugins\u002Fskills\u002Fskills\u002Fwrite-human\u002Fscripts\u002Ftest_draft_lint.py",[56,63,70,77,84,91],{"skillNo":57,"name":58,"nameZh":6,"description":59,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":60,"categories":61,"keywords":62,"publishedAt":18},"01M7DJXFGN1RJ2VEBZFT1D7YPD","worker-brief","Write the brief for a subagent, background agent or headless Claude session so it can succeed with no memory of this conversation, then lint it before dispatch. Use whenever you delegate — spawning an Agent or Task call, handing a job to a subagent, writing a prompt for another agent or session, queueing work for later, or re-briefing a worker whose first attempt missed. Covers acceptance first, inputs checked now, gotchas, write scope, authority, budget and partial returns, a separate report path, and permission to push back. For several workers on one job, use fanout-brief as well.","https:\u002F\u002Fgithub.com\u002Fmarliechorgan\u002Fclaude-setup\u002Fblob\u002Fmain\u002Fplugins\u002Fskills\u002Fskills\u002Fworker-brief\u002FSKILL.md",[],[],{"skillNo":64,"name":65,"nameZh":6,"description":66,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":67,"categories":68,"keywords":69,"publishedAt":18},"01M7DJX6NRJ0M06J9FTC9RK007","test-crew","Test a running app, tool, AI agent or plan in minutes with a crew of independent AI testers run in parallel against the real thing on this machine - a user who tries to get the job done, an attacker who tries to break in or misuse it, a strategist who asks whether it is worth using, and optionally a breaker for edge inputs - each finding tied to an input that reproduces it. Use when the user says test it, try it out, does this actually work, is it ready, sanity-check it, pentest or red-team it, UX test, user test, stress-test the plan, or before calling a build done, demoing or shipping it; also to test a chatbot or agent with a simulated user and attacker. Fast and local, not a diff review (review-changes) or a CI suite. Ships a planted-defect app to check the crew itself still catches what it should.","https:\u002F\u002Fgithub.com\u002Fmarliechorgan\u002Fclaude-setup\u002Fblob\u002Fmain\u002Fplugins\u002Fskills\u002Fskills\u002Ftest-crew\u002FSKILL.md",[],[],{"skillNo":71,"name":72,"nameZh":6,"description":73,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":74,"categories":75,"keywords":76,"publishedAt":18},"01M7DJWYGXQW414VWTNBXQ12CM","review-changes","Use this skill whenever someone asks you to review, check or look over a diff, pull request, patch, script or config change, however short, and before you call your own change ready. Reviews changed code, scripts or configuration — a diff, a pull request, the first working version of a feature, partial work being parked, a helper script whose output you are about to rely on, or several workers' changes combined. Use when asked to review, check or sanity-check a change, and on your own at those points during a build. Runs each new check against its whole input class, opens every caller and consumer the change touches, looks for other routes around a guard, compares the result with main, and reports only findings that matter, each with the input that reproduces it.","https:\u002F\u002Fgithub.com\u002Fmarliechorgan\u002Fclaude-setup\u002Fblob\u002Fmain\u002Fplugins\u002Fskills\u002Fskills\u002Freview-changes\u002FSKILL.md",[],[],{"skillNo":78,"name":79,"nameZh":6,"description":80,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":81,"categories":82,"keywords":83,"publishedAt":18},"01M7DJWQDWM25GZ1WCNR91VHNF","pr-writing","Write a pull request title and description a reviewer who never saw the conversation can assess quickly — the concrete problem, the resulting behaviour, how it was checked and what was left out. Use when drafting or opening a PR, preparing a branch for review, writing a PR for an open-source project, or rewriting a PR after its scope changed. Follows the target repository's template, branch and commit conventions. Drafting is not permission to push or open the PR.","https:\u002F\u002Fgithub.com\u002Fmarliechorgan\u002Fclaude-setup\u002Fblob\u002Fmain\u002Fplugins\u002Fskills\u002Fskills\u002Fpr-writing\u002FSKILL.md",[],[],{"skillNo":85,"name":86,"nameZh":6,"description":87,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":88,"categories":89,"keywords":90,"publishedAt":18},"01M7DJWGMD63YE1D4S3549ENJ3","oss-contribute","Take an open-source repository you have never opened to a small, reviewable pull request in one session — pick an unclaimed issue, open the clone safely, learn the project's rules, fix it test-first, write it up, and stop for the human to push. Use when the user says contribute to X, find me something to fix in X, any good first issues in X, ship a PR to X, or can we fix this bug upstream. Includes quarantining a stranger's repo so its CLAUDE.md, AGENTS.md and .claude\u002F folder cannot load as instructions or hooks. Stops at green tests and a diff; the human opens the PR.","https:\u002F\u002Fgithub.com\u002Fmarliechorgan\u002Fclaude-setup\u002Fblob\u002Fmain\u002Fplugins\u002Fskills\u002Fskills\u002Foss-contribute\u002FSKILL.md",[],[],{"skillNo":92,"name":93,"nameZh":6,"description":94,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":95,"categories":96,"keywords":97,"publishedAt":18},"01M7DJW68E30ZBRZB326BHAM9G","multi-agent-system-design","Design, debug and test applications built from AI agents — chatbots, assistants, tool-using agents and multi-agent workflows. Use when designing agent roles, tools, prompts, shared state or handoffs; when an agent misbehaves (asks for things it was already told, ignores an instruction, calls the wrong tool, says it did something that never happened, loops or stalls); when reviewing an agent architecture; or when proving an agent change works by testing the running app with a simulated user and checking what was actually saved. Covers retrieval and record selection, recovery and retries, rules that must hold, and explaining the design. For splitting the build across several Claude workers, use fanout-brief.","https:\u002F\u002Fgithub.com\u002Fmarliechorgan\u002Fclaude-setup\u002Fblob\u002Fmain\u002Fplugins\u002Fskills\u002Fskills\u002Fmulti-agent-system-design\u002FSKILL.md",[],[]]