Epic Infographics

Epic Infographics

You are going to produce an infographic image (PNG) by writing a single self-contained HTML file and rendering it with the bundled script. The three things that separate a great result from a mediocre one:

1. Varied data representations — not everything is a bar chart.
2. Executing one design language exactly — its tokens are law.
3. The check–render–review loop — the mechanical preflight must pass, and you MUST look at your own PNG and fix it before delivering. Never deliver a render you haven't looked at.

Workflow

1. Understand the brief

An infographic made for "everyone" lands with no one. Before touching any data, pin down three things:

• Audience — who is this for, and how much do they already know? (executives skimming, practitioners fluent in the jargon, general public)
• Goal — what should the graphic do: inform, persuade, impress, or drive one action? The goal decides what kind of fact can be the hero.
• Context — where will it live (feed, slide deck, print, link preview)? This feeds the canvas choice in step 5 and sets how loud the design can be.
• Still or animated — offer an animated MP4/GIF version alongside the still (step 12). Motion earns its keep where the graphic will autoplay (social feeds, stories, slides, landing pages); for print or a link preview the still alone is the right answer, so say so in the offer.

If the request already answers these, restate them in one line and move on. If not, ask the user before proceeding — one round of questions, as multiple-choice where you have a tool for it (e.g. AskUserQuestion) so answering takes one click. Only when you are running unattended and cannot ask, pick sensible answers and state those assumptions at delivery.

2. Gather the data

Collect the facts that serve the brief. Use what the user gave; research or derive the rest — never invent statistics; if values are illustrative, label them so. Gather slightly more than one graphic needs: the pitch in step 3 is only honest when there was more than one possible angle in the material.

3. Pitch the story

An infographic is an argument, not a data dump. From the data, draft 2–3 candidate story angles. Each pitch is a working headline, the hero number/fact that anchors it, and one sentence: what should the reader remember? The angles must genuinely differ — e.g. "how big it is" vs. "how fast it's changing" vs. "what it costs you" — not rewordings of one claim.

Present the pitches to the user as a multiple-choice question (AskUserQuestion where available, plain text otherwise), with your recommended angle first and marked as recommended. If you cannot ask, take your recommendation and say so at delivery. The chosen angle decides the hero element; every other fact is support or gets cut.

4. Find the visual metaphor

Read references/illustration-and-texture.md. List 2–3 physical objects the subject evokes and pick one that can carry data (a cup whose fill level is the value, a ladder whose rungs are the steps). This becomes the canvas's subject — the reason the graphic could only be about THIS topic. Type-led styles (swiss) may go abstract instead, but then scale and composition must do the identity work. The litmus test comes back at review: cover the text — is the topic still recognizable?

5. Pick the canvas

Aspect drives layout before style does: story/tall = vertical single-column flow; wide = 2–3 column zones; square = hero + 2×2 support grid is a safe default.

6. Pick the design language

Read exactly ONE file from references/design-languages/ and follow it completely — palette hexes, fonts, geometry, signature devices, do/don'ts.

• User asked for a vibe → match it (technical → blueprint, dramatic/tech → dark-glass, literary → editorial, warm/craft → retro-print, friendly → hand-drawn, nature/science/anatomy → naturalist-plate, systems/places/playful-spatial → isometric-world, places/journeys/ vistas → park-poster, how-it-works/inside-a-machine → cutaway).
• No preference → pick by subject matter, favoring the scene-native styles above.
• High slop-risk styles — swiss, corporate-clean, neo-brutalist — are flat by construction (type + shapes on a plain ground) and have repeatedly failed audience tests. Use them ONLY when the user explicitly asks for that look, and even then import scene-craft: environment, detail density, a story path (composition.md rule zero).
• Every color and font size in your HTML must come from the style file. No freelancing.

7. Choose data representations

Read references/data-vocabulary.md and pick deliberately varied forms: one hero element (big number, hero chart, or hero pictogram) plus 2–4 support elements of different types. Three bar charts in a row is a failure even if each one is correct. Prefer forms that fuse with the metaphor (liquid fill, object-as-bar, icon army) over generic charts beside it.

8. Compose and build

Read references/composition.md and pick one named composition pattern (Big Object, Bleed, Overlap stack, Diagonal drive, Editorial spread, Specimen sheet) — state it in an HTML comment. A card grid may only be a sub-zone (≤ ⅓ of the canvas), never the whole layout. Apply the tension rules: one dense zone + one empty zone, three sizes minimum, something crossing a boundary, a non-uniform background, canvas-level texture if the style calls for it. Then start from templates/skeleton.html. Rules:

• Single self-contained file: style tokens as CSS custom properties, all charts as inline SVG (recipes + math in references/charts.md), Google Fonts via <link>, no other external resources, no JS frameworks.
• One hero element; everything else visually subordinate to it. Mark it with the data-hero attribute — the preflight checker uses it to verify emphasis.
• A consistent spacing scale (the style file defines it) — no ad-hoc margins.
• Footer strip: data source (if one exists) and/or attribution.

9. Preflight the layout — mechanical gate

node scripts/check.mjs infographic.html --preset square

Before rendering a single pixel, the checker loads the page headless and measures the actual glyph geometry. It reports errors (text painted over other text, text clipped by an overflow ancestor or cut by the canvas edge, content escaping the canvas, fonts below the readable floor, more than one data-hero) and warnings (missing data-hero, weak hero dominance, tight font sizes, near-miss overlaps).

• Errors block you. Fix the HTML and re-run until the error count is zero. Never render around an error — an error at this stage is a real defect, not a formality.
• Warnings are review items: fix each one, or keep it only with a one-line justification (e.g. a deliberate 10px caption in a print titleblock).
• Ink-on-ink layering you designed on purpose (a giant translucent numeral behind a headline) is waived with data-overlap-ok on one of the two elements. Add it only for overlaps that are a composition decision, and confirm their legibility in the PNG review — every waiver is yours to defend.

10. Render

node scripts/render.mjs infographic.html infographic.png --preset square

First-time setup (once per machine): npm install && npx playwright install chromium (run from the skill directory). The script waits for fonts, so text always renders.

11. Review your own PNG — mandatory

Read the PNG file (view the image). Check ruthlessly:

• Any text overflowing, clipped, colliding, or widowed?
• Does the hero element dominate at a glance? Squint test: is there ONE focal point?
• Are chart proportions truthful (bar ratios match value ratios)?
• Contrast: is every piece of text comfortably readable?
• Crowding: does anything need more breathing room?
• Is the bottom edge awkward (half-empty, or content jammed against it)?
• Does it look like the design language, or like a generic default?

Anti-slop pass (any hit = recompose, not patch):

• The no-text squint: cover the words — can you still tell the topic?
• Could this layout hold any other dataset unchanged? (= it's a template)
• Is it rounded cards in a grid on a flat background? (= dashboard)
• Is it type + shapes on a plain ground — no scene, no environment, no story path? Flat graphic minimalism is slop even when disciplined. (= build the place; see composition.md rule zero)
• Any emoji standing in for an icon?
• Is everything evenly spaced with nothing crossing a boundary?
• Is every element between 16–40px with no giant anchor?

Fix, re-run the preflight (a fix can introduce a new collision), and re-render. Minimum one loop; repeat until the list is clean. Deliver the PNG and offer the HTML source.

12. Animate — when the user opted in

You offered animation in step 1; do this step if they said yes there, or ask for it later (MP4/GIF, social video, "make it move"). Animate the approved static file — never animate around an unreviewed still. Read references/motion.md and follow it: storyboard the build order (place → structure → data → hero → support), add CSS keyframes to the same HTML file with the fill mode motion.md prescribes, then render:

node scripts/animate.mjs infographic.html infographic.mp4 --preset square --sheet sheet.png

Requires ffmpeg on PATH (brew install ffmpeg / apt install ffmpeg). The preflight and still render are unaffected — both evaluate the animation's end state, so step 9's guarantees still hold. Review the contact sheet against the checklist in motion.md (frame 1 has a scene, the stagger is visible, the last frame matches the approved PNG), fix, re-run. Deliver the MP4 (plus a GIF via a second run with a .gif output when the destination wants one) alongside the still.

Hard rules

• No invented data. Use what the user gave, what you researched, or label values as illustrative.
• Truthful geometry. Bar lengths, arc angles, and areas are computed from values, never eyeballed. Area scales with value (a 2× value bubble has 2× area, i.e. √2× radius).
• No dual-axis charts. Two measures of different scale → two charts or index to a common base.
• Donut/pie only for part-to-whole at a glance, ≤ 5 segments, never for comparing close values.
• A static image has no tooltips — every data mark that matters carries a direct label or is readable off a labeled axis. But label selectively: the endpoint, the extreme, the hero — not a number on every point.
• Text never wears the data color — labels use the style's ink/muted tokens; identity comes from the colored mark beside them. (Exception: text set inside a colored fill picks white or ink by the fill's luminance.)
• Series colors come from the style's chart-safe slots, in their fixed order. Past 5 series, fold the tail into "Other."
• Numbers formatted for humans: 12.4M not 12400000; units always shown; large standalone numbers use proportional (not tabular) figures.
• No text overflow, no ellipsis truncation — resize, rewrap, or rewrite.
• Marks are thin, chrome is quiet (per style file): hairline gridlines if any, generous padding, saturated color reserved for data and accents.
• Never emoji as icons or illustration — draw the shape in SVG or omit.
• No framework-default colors (Tailwind/Bootstrap hexes) — palettes come from the style file, which is art-directed and validated.

Files