Typeset professional documents with Kami templates: resumes, one-pagers, white papers, letters, portfolios, and slide decks. Use when asked to 做 PDF / 排版 / 简历 / 一页纸 / PPT / slides, or to create a Kami
Install
The skill file stays in the publisher's repository.
Paste the copied command in a terminal. It installs this SKILL.md into that agent.
npx skills add tw93/Kami --skill kami -a cursor -y
Typeset professional documents with Kami templates: resumes, one-pagers, white papers, letters, portfolios, and slide decks. Use when asked to 做 PDF / 排版 / 简历 / 一页纸 / PPT / slides, or to create a Kami landing page. Not for auditing or restyling an existing product site.
name: kami
description: 'Typeset professional documents with Kami templates: resumes, one-pagers, white papers, letters, portfolios, and slide decks. Use when asked to 做 PDF / 排版 / 简历 / 一页纸 / PPT / slides, or to create a Kami landing page. Not for auditing or restyling an existing product site.'
kami · 紙
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, drafts a profile; show the draft and use it for this document only after they accept it.
python3 scripts/brand_extract.py <url>
1 · Lock the contract
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:
Current task
Mode
Contract
New document or substantial restructuring
New document
Lock the contract, write content.json with brief + content, then fill and deliver
Text replacement, translation, or factual correction in an existing artifact
Content-only
Preserve CSS and layout unless the new copy proves a fit defect
User supplies a render or screenshot and rejects how it looks
Visual repair
Treat the render as the brief; follow «Feedback protocol»
Standalone generated illustration, cover, social card, or redraw
Generated asset
Lock the semantic brief before pixels (references/diagrams.md section 11); preserve accepted parts across iterations
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.
Language
User language
HTML templates
Slides (PDF default)
Slides (PPTX, only when an editable deck is requested)
Slides (Marp)
Chinese (primary)
*.html
slides-weasy.html
slides.py
slides-marp.md
English
*-en.html
slides-weasy-en.html
slides-en.py
slides-marp-en.md
Japanese (best-effort)
*.html, JP Mincho first
slides-weasy.html
slides.py
slides-marp.md
Korean
*-ko.html
slides-weasy-ko.html
slides-en.py only if PPTX is required
slides-marp.md
Other (best-effort)
CJK or EN path by script coverage
slides-weasy.html or -en
slides.py / slides-en.py
slides-marp.md or -en
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:
Signal
Document
1 page for investors, recruiters, or an executive summary
one-pager
1 page of formal correspondence (sales, hiring, resignation, memo)
letter
Exactly 2 pages of career narrative with project bullets
resume
3-6 visual-heavy pages showcasing projects
portfolio
6-15 pages of sustained argument, low visual density
long-doc
Presentation flow, one assertion per slide
slides
Metrics dashboard plus thesis plus price or risk view
equity-report
Version-by-version release facts
changelog
Product showcase, pricing, screenshots, FAQ for a browser
landing-page
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?).
Changelog vs. release notes: the changelog template is a styled document. GitHub release notes are a separate deliverable; use /write.
Landing page: screen-first HTML, no PDF. Fill the {{PLACEHOLDER}} values and comment blocks and save a ready-to-serve file. 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.
Diagrams inside a document come from 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.
Charts from data (pick and embed them without asking) are generated, never drawn by hand: write the numbers as a spec and run python3 scripts/build.py --chart spec.json -o chart.svg (references/diagrams.md section 9).
Slides: read references/deck-preflight.md before drafting.
2 · Sources, materials, and content
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.
3 · Read the right amount
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.
Task
Read
Content-only change
CHEATSHEET.md
Layout tweak within spec
CHEATSHEET.md + the template
New document
the template + CHEATSHEET.md + references/schemas/<type>.json + the references/writing.md sections for that type («Per-document strategies», «Quality bars by document type»); open references/design.md only for the component or rule you are about to change
+ references/deck-preflight.md and the references/design.md section 8 «Deck Recipe» it points to
Diagram
references/diagrams.md; references/mermaid.md when the source is Mermaid text
Company, product, market, or branded subject
references/writing.md sections 5-6
Rendering bug, font issue, overflow
references/production.md Part 4 (and the design spec if CSS is the cause)
Reviewing an AI-generated draft
references/anti-patterns.md
4 · Fill the template
Copy the template into the working directory; never write HTML from scratch. CSS stays untouched during a content fill. A layout change stays within spec, and a real style change moves references/design.md and the sibling templates together, never one file.
Meet the quality bar for the type in 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.
Do not invent metrics, financial data, or statistics, and keep any figure you compute from the supplied ones next to its inputs or marked as derived; do not describe stock images as placeholders; do not pad slots (three real projects do not become five). Avoid the patterns in references/anti-patterns.md.
Tables separate rows with whitespace first and neutral hairlines second; decoration earns its place only by encoding data, state, grouping, or a relationship (references/design.md «Table (kami-table)» and «Subtractive rule»).
Resume: run the recruiter pass in references/resume-writing.md before building.
Mathematics is strict LaTeX: author formulas only as inline \( ... \) 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.
Code blocks with 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:
CN
EN
Rule
{{作者}}
{{AUTHOR}}
Resume, letter, portfolio: the person's name from the document. Others: leave it, the build infers it from git config user.name, then KAMI_AUTHOR
{{摘要}}
{{DESCRIPTION}}
One sentence (≤150 chars) from the first two paragraphs
{{关键词}}
{{KEYWORDS}}
3-5 keywords from the title and section headings
{{文档标题}} etc.
{{DOC_TITLE}} etc.
From the H1 or .header .title
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.
5 · Deliver
One command finishes a document. Run it from the document's directory, fix every ERROR, and re-run until it reports READY:
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) 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:
Landing pages and any browser surface: --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.
Editable PPTX: python3 scripts/build.py --verify slides (or slides-en) and --check-rhythm.
Missing dependencies or fonts: 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:
The path of every deliverable, in every promised format.
The --deliver verdict and page count against the contract.
Every [DATA NEEDED] gap and every assumption you made, listed.
The visual verdict stated honestly: when page images could not be inspected, say "build verified, visuals unconfirmed", not "done".
Feedback protocol
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.
Name the concrete defect in one sentence: page or surface, viewport or state, and whether it is density, hierarchy, alignment, type, color, cropping, or text fit.
Lock the repair boundary: 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.
Make the smallest content, geometry, spacing, typography, crop, or token change that fixes the defect. Never hide a content problem by shrinking type first.
Verify the affected matrix rather than one screenshot:
PDF: target page, neighboring pages, total page count, font result, and every locale or template variant reached by a shared token.
Screen: the breakpoint, tablet, and baseline matrix in 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.
PPTX: editable source plus a rendered PDF or opened-deck inspection.
Generated asset: target slot at its smallest display size plus sibling assets in the same deliverable.
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.
When not to use this skill
The user explicitly wants Material, Fluent, or a Tailwind default: a different design language.
Dark, cyberpunk, or futurist aesthetics (this is deliberately anti-future).
Saturated multi-color palettes (this has one accent).
Cartoon, animation, or illustration style (this is editorial).
Dynamic web app UI (this is for print and static documents).
How do I install kami?
Download plugins/kami/skills/kami/SKILL.md from the publisher, or copy the path and open the repository. PlugMCP lists the skill and does not host the file.
Where does kami live?
The publisher keeps it at https://github.com/tw93/Kami.
Which agents can use kami?
Claude, Cursor, and Codex can read a SKILL.md you add from the publisher's repository.