Turn a structured document (an /ingest reflection, role report, weekly-learnings digest, deck outline, research note) into a beautiful, self-contained HTML infographic — a visual one-pager — with brand-aware theme selection. Use when the user asks to "make an infographic," "visualize this note/report," "turn this into a one-pager/poster," or accepts the post-ingest "generate a visual infographic?" offer.
설치
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Turn a structured document (an /ingest reflection, role report, weekly-learnings digest, deck outline, research note) into a beautiful, self-contained HTML infographic — a visual one-pager — with brand-aware theme selection. Use when the user asks to "make an infographic," "visualize this note/report," "turn this into a one-pager/poster," or accepts the post-ingest "generate a visual infographic?" offer.
Infographic Builder
Take a finished document and distill it into a single, share-ready HTML infographic. Not a slide deck, not a dashboard — a one-page visual argument: hero thesis, a few hero stats, the narrative arc, the key contrast visualized, the takeaways as cards.
The output is one self-contained .html file (Tailwind via CDN + Google Fonts, no build step) that opens anywhere and can be rasterized to PNG/PDF for sharing.
When to use
A document already exists and is worth a visual summary (ingests, role reports, weekly learnings, research notes, strategy memos).
The user says "infographic," "visualize this," "one-pager," "poster," or takes the /ingest post-step offer.
Do not use this to build dashboards (always-on metric views → aios/data-presentation) or multi-page decks.
Process
1. Read the source, extract the spine
Read the full document. Pull: the one-line thesis, the 2–4 hardest stats/facts, any narrative sequence (steps/timeline), any tension or dichotomy (X vs Y), the takeaways/lessons, the best pull-quote, and the sources.
Fact discipline (non-negotiable). Use ONLY facts present in the source. Never invent metrics, dates, impression counts, percentages, or compositional details to make a panel look fuller. If the source doesn't say it, it doesn't go on the infographic. (A blank-but-true panel beats a full-but-fabricated one. Common LLM failure: inserting plausible-sounding metrics like "~12M impressions in 72 hours" or invented chemical/material names — plausible, narratively useful, not in the source. Skip the stat or replace with a true qualitative descriptor. See feedback_no_fabricated_specifics.) When a number is striking, attribute it the way the source does.
2. Map to the Infographic IA
Distill into the seven-section taxonomy. Not every section is mandatory — include only what the content earns. Order is suggestive; what matters is the hierarchy of attention.
#
Section
Purpose
Shape
1
Hero
Argument in 3 seconds.
5–9-word title (noun phrase, the thesis) + a one-line subtitle that states what the page proves.
2
Stat callouts (1–3)
The hardest, most arresting facts.
One "before/positive" stat ↔ one "after/negative" stat, paired and source-attributed. If the source has only one hero number, use one card, not two.
3
Narrative arc → timeline
The sequence, in order.
3–6 steps; each = short label + one-line mechanism + one-line outcome. Step grid or tabbed cascade. Skip if no sequence.
4
Contrasting forces → paired cards
The tension visualized.
Two parallel 3-card stacks (X-layer vs Y-layer, promise vs reality, what-worked vs what-failed). Items must be semantically paired, not just thematically. Skip if no real dichotomy.
5
Lessons
What generalizes.
Exactly 3 cards. One-phrase title + 2–3 lines of explanation. The structural lessons — not a recap.
6
Key Contrast (the "wow")
The one panel people screenshot.
ONE viz of the single quantitative or structural asymmetry the whole piece pivots on. If the source has no single pivot, cut this section — do not fabricate one.
7
Quote + sources
Voice + traceability.
One pull-quote with attribution + source list (URL/path). Every striking number elsewhere on the page traces back here.
3. Select the theme (matching hierarchy)
In priority order — and always tell the user which theme you picked and why, and offer a one-line override:
Explicit reference — the user named a brand, dropped a design.md, or pointed at one → use it verbatim.
Brand-first — if the content maps to a venture (frontmatter venture:, wiki-links to a venture, tags), use that venture's design system: vault/00 - notes/context/ventures/{venture}/design.md. (Per the AIOS rule: read the design system, never hand-pick fonts/colors.)
Fallback library — otherwise pick (or combine) the best-matching DESIGN.md from VoltAgent/awesome-design-md by mood/domain. This is the default when there is no input or design reference. Fetch at runtime:
(<slug> is the lowercase brand name; confirm it exists first — slugs vary, e.g. wired resolves, some don't). Combining two is allowed when it sharpens — pick one as the typography spine and the other as the surface/palette.
Local alternative: anthropic/theme-factory themes are also available if no web fetch is desired.
Mood → brand starting points (suggestive, not authoritative — read the actual DESIGN.md to confirm):
Content tone / domain
Try first
Why
Investigative, editorial, journalism
wired
Print-magazine gravitas. Note: light, not dark.
Dark, precise, technical / SaaS
vercel, supabase (dark variant), linear
Modern dark editorial.
Warm, trustworthy, AI / research
claude, cohere, mistral
Approachable; Anthropic-adjacent.
Fintech, money, identity
stripe, coinbase, wise
Restrained, institutional.
Consumer playful
spotify, airbnb, notion
Color-forward, friendly.
Automotive / luxury / kinetic
tesla, ferrari, bmw
Bold, sculptural.
4. Render
One self-contained HTML file. Tailwind (CDN) + Google Fonts. Apply the chosen design system's actual tokens — color palette + roles, typography hierarchy, component stylings, depth/elevation. Write the markup directly (do not fill a JS template-string — that leaks ${placeholders} under pressure). Responsive (mobile → desktop). Light, purposeful interactivity only (a tabbed cascade for a timeline is good; gratuitous motion is not). For visual craft follow anthropic/frontend-design + anthropic/canvas-design; for which-viz-for-which-data follow aios/data-presentation.
Render checklist:
Self-contained single .html (no build step). Tailwind via CDN + Google Fonts.
Design tokens applied from the chosen DESIGN.md, verbatim — exact hex values, exact font families, exact weights, exact spacing scale. No drift.
Markup generated directly — no JS template-string fill.
Dark/light toggle, unless explicitly out of scope. Architecture: CSS custom properties for surface/text tokens (:root = one mode, [data-theme="..."] = the other). Brand colors (primary, accent) stay constant — they're identity. Tailwind colors point at the vars (rgb(var(--c-surface) / <alpha-value>)) so existing classes adapt without markup churn. Include a header toggle button + a tiny inline FOUC-prevention script in <head> that reads localStorage + prefers-color-schemebefore render. Default to whichever mode the chosen DESIGN.md leans toward; let the user flip. ~30 lines of CSS + 15 of JS — not overweight.
Selection color, scrollbar styling, focus rings — small details that make it feel finished.
Light interactivity only when it earns its place.
All facts traceable to the source. Sources section at the bottom with URLs/paths.
No ${placeholders} in output. Grep for $ before declaring it done.
External scripts: Subresource Integrity (SRI). For shareable / embedded / hosted use, every <script src="https://..."> and CDN <link href="https://..."> MUST carry integrity="sha384-…" + crossorigin="anonymous" against a pinned version (jsdelivr/unpkg with an exact version + hash). For Tailwind Play CDN (cdn.tailwindcss.com): no stable hash exists by design (JIT-compiled), and Tailwind itself states it is not for production — acceptable only for local preview artifacts the user opens from disk. For anything hosted/embedded/attached, swap Tailwind Play CDN for either (a) a pinned tailwindcss@X.Y.Z build from jsdelivr with SRI, or (b) inline the compiled CSS. Default the local-preview build to Play CDN with a comment noting the production swap; flip the moment the artifact leaves file://.
5. Output + optional export
Two destinations, split by weight. The HTML is the canonical artifact — lightweight, self-contained, version-controlled — and goes in the vault. Raster exports are ephemeral views and go to ~/Downloads, never the vault.
HTML → vault. Default path + naming:
03 - export/infographics/{YYYY-MM-DD}-{slug}.html
{YYYY-MM-DD} = the render date (not the source-doc date), so a re-render lands as a new file rather than silently overwriting.
{slug} = lowercase-kebab-case derived from the source filename or the one-line thesis (strip dates, articles, punctuation; max ~6 words).
Mirrors the vault's date-prefixed convention used in reflections/ingests/ and other dated artifacts.
Raster exports → ~/Downloads (offer these; don't auto-run). Never the vault:
PNG — ~/Downloads/{YYYY-MM-DD}-{slug}.png — Chrome headless --screenshot per the rasterize recipe (native viewport + --force-device-scale-factor=2).
PDF — ~/Downloads/{YYYY-MM-DD}-{slug}.pdf — mcp__pdf-generator__html_to_pdf.
Why the split: the HTML regenerates any raster on demand, so a committed PNG/PDF is pure bloat — a 1 MB+ binary that re-downloads on every clone and re-syncs on every push, for a view you can rebuild in one command. HTML in the vault; rasters in ~/Downloads. (If the caller explicitly names a different output dir, honor it — but default rasters outside the git-synced vault.)
Then surface the file to the user (e.g. SendUserFile for the raster, or the HTML path).
Composition
This skill orchestrates existing ones — it doesn't reinvent them:
