本文へ移動
cccskills
無料GitHub で公開

editorial-illustrations

Generate meaning-carrying editorial data-illustrations in the monotykamary / Linear aesthetic (near-black grayscale, Inter display + mono labels, hairline framed figures) with a single coral accent. This is a GENERATIVE GUIDE, not a template gallery: it teaches the "claim -> geometry" method so any session can invent the RIGHT figure for what a post actually argues, then render it as a self-contained, theme-aware, reduced-motion-safe HTML/SVG figure. Use when writing/illustrating a blog post, essay, spec, or slide and you want a chart, diagram, or illustration that carries a thesis (not decoration) in the house style. Triggers: "make a chart/diagram/illustration for this", "visualize this argument", "monotykamary/Linear style figure", "editorial graphic".

インストール方法を見る

含まれるファイル(5)

  • SKILL.md9.5 KB
  • assets/gallery.html24.2 KB
  • references/design-system.md9.5 KB
  • references/elements.md8.2 KB
  • references/worked-examples.md13.6 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Editorial illustrations — claim → geometry

The look you want (monotykamary.com, linear.app/now) is not a chart library. It is a discipline: every figure is a single argument rendered as geometry, framed in a hairline box, captioned in mono uppercase with the meaning, drawn in grayscale with exactly one accent. Get the discipline right and any shape reads as "house." Copy templates blindly and it reads as slop.

Do not reach for a fixed set of charts. Reach for the method below and build the figure this specific sentence needs. The six worked examples in references/worked-examples.md are proof the method generalizes, not a menu.

The core move

Take the one sentence the figure must prove. Find that sentence's structure. Render the structure so the reader believes the sentence before reading it.

The caption is the thesis, written first. If you can't state the figure's claim in one sentence, you don't have a figure yet, you have decoration. Stop.

Method (run this every time)

  1. Write the claim. One declarative sentence. This becomes the <figcaption> verbatim. Example: "Nothing is billable until work crosses the delivery boundary."

  2. Name the claim's structure. Which of these is the sentence really about? Match to a geometric primitive:

    If the claim is about…Render it as…Atom (§5)
    a spectrum / more-vs-less / a gradientordered cells, rising fill5a
    ranked magnitudes / shares / a comparisonhorizontal bars, real axis5b
    a sequence with a tipping pointstage grid + traveling dot5a + rail
    mutual causation / a feedback relationshiptwo coupled orbits + bridge5d
    a branch / decision under uncertaintynode-and-branch SVG tree5c
    one thing changing form / a data pipelinestage rail, a glyph per stage5e
    accumulation toward a targetstacked / filling bar5a/5b
    a boundary / inside-vs-outsidea framed region + a crossingcompose
    a distribution / where mass sitsdot-strip or small-multiplescompose

    If nothing matches, compose a new primitive from the atoms in references/design-system.md §5. The novelty is expected and welcome; the tokens and restraint are what must stay constant.

  3. Locate the point. Which single element proves the claim? That one gets --accent (coral). Everything else stays grayscale. One accent, maybe two. No point → the figure has no thesis → go back to step 1.

  4. Draw in the system. Apply references/design-system.md: tokens, the typography law (display = Inter 480–540 tight; everything else mono, tiny, UPPERCASE labels), the framed-figure chrome, 720-wide SVG viewBox, hairlines only, no chartjunk.

  5. Add motion only if it demonstrates the claim. A dot that pauses before a boundary; a pulse crossing between two loops; a signal falling down a branch. 2–6s, ease-in-out, and always killed under prefers-reduced-motion. If motion doesn't show the argument, omit it. Static is fine and often better.

  6. Caption it. Mono, 8px, uppercase, above-the-rule. State the meaning, not the data source alone. ("…because X," not "Figure 3.")

The aesthetic in one breath

  • Field: matches the host — near-black --ground:#080807 on a dark page, white/ off-white on a light page (both token sets in design-system.md). Framed figures on a faint surface; backdrop-filter: blur(16px) on dark, a hairline + faint shadow on light.
  • Type: Inter for the rare big display; mono for everything inside a figure, 7–11px, labels/captions UPPERCASE. Never a second sans, never a bold display.
  • Color: grayscale ramp (--ink → --quiet) + one accent = Claude-coral #D97757, on the point only. (Set --accent: var(--ink) for monotykamary-faithful pure grayscale.)
  • Structure: 1px hairlines, true-circle nodes, dashed = provisional/failure, no shadows, no gradients except a single axis-direction hairline.
  • Density: small type, tight cells, edge-to-edge. Restraint is the style.

Full spec + copy-paste token blocks (dark AND light) + primitive CSS: references/design-system.md. Worked examples with the claim→geometry reasoning: references/worked-examples.md. Non-chart editorial elements (metric strip, status aside, kicker+sigil, numbered chapter index, prose type, inline cards): references/elements.md. Live rendered proof (coral house style, toggleable): assets/gallery.html.

The figure follows the host page's theme (ALWAYS)

A figure is a guest on someone's page. It must match that page's light/dark, not impose its own. A dark exhibit panel dropped into a light article reads as a foreign object (learned the hard way, 2026-07-22).

  • Look at the host first. Light page → light figure. Dark page → dark figure. a host site may default light and carry its own theme toggle (e.g. Charter serif heads + Public Sans body, blue accent) that stamps data-theme="light"|"dark" on <html> — figures there must follow data-theme, defaulting light.
  • If the host has an explicit theme system, key the dark override off IT, not the media query. Use :root[data-theme="dark"] .yourfig { … }, NOT @media (prefers-color-scheme: dark). Real bug (2026-07-22): a figure keyed on the media query rendered dark on the light blog because the viewer's OS was dark and the page theme was light. The page's data-theme is the source of truth; the media query only applies on hosts that have no toggle. On a toggled host, match the site's own dark tokens (ground #171b22, line #262c36, ink #f1f5f9) so the figure's dark panel sits in the site's dark, not a foreign near-black.
  • Build token-based so it adapts. Define the palette on the figure's own scope, default = the host's default theme, plus a :root[data-theme="dark"] .fig override (and a @media (prefers-color-scheme: dark) override too ONLY if the host has no toggle). Costs nothing, future-proofs the figure.
  • Coral is AA-safe per ground. On white use --accent:#c15f3c for text/strokes (keep #D97757 only for large node fills); on near-black use #D97757 directly.
  • The claim→geometry method and the restraint rules are theme-independent. Only the palette tokens swap. Grayscale-ramp + one coral accent holds on both grounds.

Output contract

  • One self-contained <figure> per illustration: inline <style> scoped to the figure (or shared token :root), inline SVG, no external requests.
  • Theme = host's (see above): light default when embedded in a light page, but always token-based with prefers-color-scheme + :root[data-theme] overrides, and reduced-motion-safe. Both are house rules, not optional.
  • For an Artifact: no external fonts/CDN (CSP blocks them); inline @font-face data-URIs or fall back to system-ui/ui-monospace.
  • For blog posts on your own site: emit the figure HTML/CSS into the post body (markdown passes raw HTML through verbatim via build.py). The blog is light, so the light tokens apply; keep the dark override for portability. Rebuild with python3 build.py; the built page uses absolute asset paths so preview it via a local server or a self-contained standalone copy, never a bare file:// open (that renders the post unstyled). See a local self-contained HTML preview.

Verify (mandatory — this is a UI/UX output)

Per the Visual Verification rule, a DOM check is not enough. Render it and look.

  1. Open the figure in browser-harness, screenshot it (evidence_shot).
  2. Read the image against the claim: does the geometry make the sentence obvious? Is coral on the point and nowhere else? Any clipped mono text, misaligned SVG labels, off-by-one bar math, broken responsive stack at ≤720px?
  3. Fix what you see, re-shot to prove it. Check the light variant too.
  4. Confirm motion stops under reduced-motion.

Anti-slop checklist (reject the figure if any fail)

  • The figcaption is a real claim, stated in one sentence.
  • The geometry encodes that claim's structure (not a generic chart pinned on).
  • Exactly one accent element (coral), on the point.
  • Every in-figure label/value/caption is mono + tiny; labels UPPERCASE.
  • Hairlines only; no shadows, no gradients (bar-axis hairline excepted), no rounded rectangles.
  • Gridlines present only if the figure is genuinely quantitative.
  • Motion (if any) demonstrates the claim and dies under reduced-motion.
  • Renders correctly in both dark and light, and stacks cleanly on mobile.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Deep strategic research engine — decomposes questions into parallel research threads, spawns multiple agents, and synthesizes into actionable strategic analysis

日本語の概要は準備中です。原文の説明を表示しています。

huytieu/COG-second-brain1,2762026年10月2日 更新

braindump

無料

Quick capture of raw thoughts with intelligent domain classification and competitive intelligence extraction

日本語の概要は準備中です。原文の説明を表示しています。

huytieu/COG-second-brain1,2762026年10月2日 更新

Run one task through the V-model verification loop: CP-2 plan → CP-3 build → CP-3v component verify → CP-4 integration verify (full lane) → CP-5 acceptance. The worker never grades its own homework; evidence rows trace back to AC-n. Opt-in: invoke with /closed-loop or by asking for the closed loop, proper verification, or an evidence trail. Ordinary work does not run this.

日本語の概要は準備中です。原文の説明を表示しています。

huytieu/COG-second-brain1,2762026年10月2日 更新

Deep-dive 7-day analysis across all data sources for weekly reviews, board prep, and strategic planning

日本語の概要は準備中です。原文の説明を表示しています。

huytieu/COG-second-brain1,2762026年10月2日 更新

Autonomous content pipeline - scout announcements in your field, triage by trend momentum and personal angle, produce posts/blogs/videos in your voice with ledger-based dedup, hard volume caps, and screenshot-verified publishing

日本語の概要は準備中です。原文の説明を表示しています。

huytieu/COG-second-brain1,2762026年10月2日 更新

Create user stories with duplicate checking across any project tracker (Linear, GitHub Issues, Jira)

日本語の概要は準備中です。原文の説明を表示しています。

huytieu/COG-second-brain1,2762026年10月2日 更新

huytieu のスキルをすべて見る

このスキルの問題を報告する