- name
SKILL.md
स्किल फ़ाइल पब्लिशर की रिपॉज़िटरी में ही रहती है।
कॉपी की गई कमांड टर्मिनल में पेस्ट करें। यह इस SKILL.md को उस एजेंट में इंस्टॉल करती है।
npx skills add tw93/Kami --skill kami -a cursor -yकॉपी की गई कमांड टर्मिनल में पेस्ट करें। यह इस SKILL.md को उस एजेंट में इंस्टॉल करती है।
npx skills add tw93/Kami --skill kami -a claude-code -yकॉपी की गई कमांड टर्मिनल में पेस्ट करें। यह इस SKILL.md को उस एजेंट में इंस्टॉल करती है।
npx skills add tw93/Kami --skill kami -a codex -yकॉपी की गई कमांड टर्मिनल में पेस्ट करें। यह इस SKILL.md को उस एजेंट में इंस्टॉल करती है।
npx skills add tw93/Kami --skill kami -a grok -yकॉपी किया गया संदेश Grok Bot चैट में भेजें। Bot यह SKILL.md GitHub से लाता है और उसे उन स्किल में जोड़ देता है जो आपके सभी Bot साझा करते हैं।
Install this skill: https://github.com/tw93/Kami/tree/HEAD/skills/kamiपब्लिशर से
紙 · かみ means paper in Japanese.
Good content deserves good paper. Use Kami’s templates and layout rules to create documents and landing pages with serif fonts, warm backgrounds, and ink-blue accents.
Part of Kaku · Waza · Kami - Kaku writes code, Waza drills habits, Kami delivers documents.
Commands and paths below are relative to this skill's directory. Keep the document in the user's working directory and call the scripts from there by their full path (python3 <skill>/scripts/build.py ...); relative file arguments resolve from your working directory first.
Update check (non-blocking). At the start of a task, run bash scripts/check-update.sh. It keeps a local daily marker and, at most once a day, reads Kami's latest published GitHub Release; it uploads no document or task content and stays silent when offline or sandboxed. When it prints a line about a newer Kami, relay it and continue.
Brand profile. If ~/.config/kami/brand.md (or the legacy ~/.kami/brand.md) exists, read references/brand-profile.md and apply it. Precedence: explicit prompt > editorial judgment > habit notes > frontmatter defaults > built-in defaults. The profile fills gaps silently and never overrides the current conversation. When the user names another project as the visual reference, run the style scan in the same file. When they give only a brand website, python3 scripts/brand_extract.py <url> drafts a profile; show the draft and use it for this document only after they accept it.
Before creating or changing an output, lock six things: language, template, output formats, page or length target, visual acceptance, and the verification command. Infer them from the request, then state the contract to the user in one short line in their language (template, length, formats, the figures you plan, any missing material) and continue without waiting. If they push back, adjust.
Question budget. Infer first. Ask at most once, in one compact question with at most two parts, and only when two or more of purpose, audience, hard constraints, or success are genuinely unresolvable, or two templates genuinely both fit. Otherwise take the smallest reasonable assumption and report it at the end.
Route by the artifact's current state:
Use the nearest existing template and check. Do not add a template, shared CSS layer, dependency, script flag, or optional mode unless the request cannot be met without it.
Japanese, Korean, and other languages need a visual pass on the page images before shipping. Reference docs are shared English specs.
When two cells both fit, decide from the signals before asking:
Ask only in cases like "1.5 pages of career story, heavy visuals" (resume or portfolio?) or "2 pages of exec summary with metric tiles" (one-pager or equity-report?).
/write.{{PLACEHOLDER}} values and comment blocks and save a ready-to-serve file. The accent follows the product: use its own main color, ask the user to pick when it has none, and keep ink-blue only as the fallback (references/design.md Section 11 «Accent on a product site»). Companion files (landing-page-*.example) and the production site mode (docs, help, releases, legal pages, or more than two locales) live in references/design.md Section 11. Keep project-specific release artifacts, payment providers, appcast rules, and private local paths out of Kami.assets/diagrams/ (eighteen types), not a template. references/diagrams.md section 1 picks the type; extract the <svg> block into a <figure>. The same file owns architecture boards, diagrams maintained in the user's repo, and generated raster illustrations.python3 scripts/build.py --chart spec.json -o chart.svg (references/diagrams.md section 9).references/deck-preflight.md before drafting.<section class="card"> per 1242x1660 image and one paper for the set, set on <body> (references/design.md Section 13). The cover headline is the post title and the cover carries a real photo or screenshot; the author's text stays as written, and a short last line is fixed with layout, not by rewording. --deliver writes each card at its pixel size.Source check. When the document names a company, product, person, release date, version, funding round, metric, market fact, or technical spec, work from primary sources, note source and date for the facts that drive it, and ask when sources conflict (references/writing.md «5. Sources before phrasing»). Skip it for personal drafts where the user supplied everything.
Material check. A branded subject needs what makes it recognizable: logo, product image or current UI screenshot, and brand colors when the document is branded (references/writing.md «6. Materials serve recognition»). When the request names no logo but the brand profile has one, fill the commented .brand-logo slot in one-pager / portfolio / slides-weasy (expand ~ to an absolute path; leave the slot commented if the file is missing). An explicit logo in the current request always wins.
Gap report. Everything the template needs that the input lacks, missing facts and missing materials alike, goes into one compact table shown once, with one question only when a missing item would change the deliverable. Inside the document, mark a gap as [DATA NEEDED: what]. Never fill a gap with a guess, a stock image, an approximated logo, or an invented value.
Distill. Unless the content already maps one-to-one onto the template sections or the user said to use it as-is, distill it. Every factual claim, number, date, name, source, and material reference from the input must land in a template section or in the gap report.
Content IR (new documents). Write the distilled content and the locked contract to content.json next to the working HTML before filling. Read references/schemas/<type>.json first: its $comment notes carry the per-field quality bar. The envelope is strict: only type, lang (cn, en, ko, zh-TW, ...), brief, and content.
{
"type": "resume",
"lang": "cn",
"brief": {
"audience": "Hiring manager for a senior product role",
"job": "Earn an interview by proving scope and outcomes",
"template": "resume",
"formats": ["html", "pdf"],
"page_target": 2,
"narrative": "Scope first, then evidence, then fit",
"required_facts": ["team size", "measured outcomes"],
"required_assets": [],
"acceptance_checks": ["two pages", "all atomic facts survive"],
"preserve": [],
"explicit_deviations": []
},
"content": { ... }
}
type is a schema name in references/schemas/ (one-pager, letter, resume, long-doc, portfolio, slides, equity-report, changelog, landing-page). For a visual repair, add target, evidence, and preserve to brief. Validate before layout with python3 scripts/build.py --check-content content.json and fix errors by fixing the content (or asking for the fact), never by loosening structure. Short atomic values (names, metrics, dates, bullets) must survive into the document verbatim; values over 80 characters may be rephrased.
Read the lowest tier that covers the task. The templates already carry every token and component, so a new document does not need the full design spec.
references/design.md and the sibling templates together, never one file.references/writing.md «Quality bars by document type»: a resume bullet needs action, honest scope, and an observable result; an equity report needs variant perception and quantified catalysts; slides need assertion titles. Structure is necessary, not sufficient.references/anti-patterns.md.references/design.md «Table (kami-table)» and «Subtractive rule»).references/resume-writing.md before building.\( ... \) or display \[ ... \] in rendered body text, never as Unicode approximations, ASCII fractions, screenshots, or raw TeX on the page. --deliver renders them to MathJax SVG in place; it needs Node.js 20 or 22+ and bash scripts/ensure_mathjax.sh once.class="language-*" are highlighted when optional Pygments is installed; without it they render monochrome.Fill the four metadata placeholders in <head>; WeasyPrint writes them into the PDF:
Keep <meta name="generator" content="Kami"> as is.
Output formats. Do not ask; decide from context. Any document: HTML + PDF (PDF is the deliverable, HTML the source). Slides: HTML + PDF, plus PPTX only when an editable deck is requested. "分享 / 发朋友圈 / share / post / preview": add PNG. "嵌入 / 插图 / embed in another doc": PNG only. Landing page: the static HTML file. Word or .docx for long-doc, proposal, or equity-report: the PDF plus a DOCX through references/production.md Part 5. An explicit format request overrides all of this.
One command finishes a document. Run it from the document's directory, fix every ERROR, and re-run until it reports READY:
python3 scripts/build.py --deliver filled.html content.json
It checks the HTML (placeholders, strict math rendered in place, Markdown residue, generated charts against their data, template style, content coverage), renders the PDF next to it, then checks the PDF (page contract, resume balance, the font that actually drew CJK text, residue, density, orphans; for social cards one page per card, card fit and card orphans) and exports page images. WARN lines point at where to look, not at failures. GAP lines are the [DATA NEEDED] markers you must report.
Then do the part no script can: open every page image and walk the printed checklist. One hit means a whole-document sweep for that class of issue. Fix the three most visible defects per round, re-run --deliver, and stop after three rounds; list what is still wrong rather than looping. If the host cannot view images, send the image paths and the checklist to the user instead of skipping the pass. A missing CJK serif shows no fallback boxes, only a heavier, flatter page, which is why the font gate exists; never call a CJK document visually verified without it.
Other surfaces:
--deliver checks the HTML only. Screenshot the responsive matrix in every shipped locale before calling it done (references/design.md «Responsive screenshot verification»); do not wait for the user to ask about mobile.python3 scripts/build.py --verify slides (or slides-en) and --check-rhythm.python3 scripts/build.py --doctor reports what is installed; bash scripts/ensure-fonts.sh recovers CJK fonts (the commercial TsangerJinKai02 never ships inside the skill package). Render failures and their fixes: references/production.md Part 4.Fresh review. After READY and the image pass, review once from the contract rather than from your own rationale: reread brief, the coverage result, and the page images; check every acceptance item and preserve boundary; fix P0/P1 findings, each tied to the page or element that proves it. Use an isolated reviewer when the host has one.
Definition of done. The closing message gives:
--deliver verdict and page count against the contract.[DATA NEEDED] gap and every assumption you made, listed.When the user gives visual feedback ("looks off", "太挤了", "not elegant"), inspect the current render before asking them to choose a value. The render is the evidence; the user's negative label is the acceptance signal.
target is allowed to change; preserve names the adjacent pages, sections, content, and shared tokens that must stay stable. Ask only when two plausible targets would produce materially different artifacts.references/design.md «Responsive screenshot verification»; every shipped locale; affected default, focus/selected, loading, empty/error, and transition states only when the surface has them.Never say "I'll adjust the spacing" without naming the exact property and its new value.
Escalate after two rounds. If the same element is still not approved after two adjustment rounds, stop nudging values and produce one comparison artifact: the current state plus 2-3 labeled variants (A/B/C) of the same content in the product's actual frame and background, changing only the compared property. For choices with no objective criterion (typeface, accent color, logo), start there: a specimen sheet of up to 5 labeled half-page blocks with identical title-plus-paragraph content. After the pick, apply it everywhere and rebuild what it touches in the same round.
Typeset professional documents with Kami templates: resumes, one-pagers, white papers, letters, portfolios, slide decks, and social media cards. Use when asked to 做 PDF / 排版 / 简历 / 一页纸 / PPT / slides
---name: kamidescription: Typeset professional documents with Kami templates: resumes, one-pagers, white papers, letters, portfolios…---# skills/kami/SKILL.mdओवरव्यू
Typeset professional documents with Kami templates: resumes, one-pagers, white papers, letters, portfolios, slide decks, and social media cards. Use when asked to 做 PDF / 排版 / 简历 / 一页纸 / PPT / slides / 小红书配图, or to create a Kami landing page. Not for auditing or restyling an existing product site.
FAQ
पब्लिशर से skills/kami/SKILL.md डाउनलोड करें, या पाथ कॉपी करके रिपॉज़िटरी खोलें। PlugMCP स्किल को सूचीबद्ध करता है, फ़ाइल होस्ट नहीं करता।
पब्लिशर इसे https://github.com/tw93/Kami पर रखता है।
Claude, Cursor, Codex और Grok Build वह SKILL.md पढ़ सकते हैं जिसे आप पब्लिशर की रिपॉज़िटरी से जोड़ते हैं। जब आप चैट में स्किल के GitHub फ़ोल्डर का लिंक भेजते हैं, तो Grok Bot उसे इंस्टॉल कर देता है।