Slide Creator

Slide Creator — HTML → PDF Presentations

Build slide decks as HTML, export to pixel-perfect 16:9 PDF via headless Chromium.

Output format: PDF — not Microsoft PowerPoint (.pptx). The PDF preserves exact layout, fonts, and colors across all devices.

Why HTML → PDF

• CSS layout (Grid/Flexbox) is far more flexible than any PPT editor
• Full web typography, gradients, SVG, animations (print degrades gracefully)
• Git-friendly, reproducible, scriptable
• One command → PDF with exact 16:9 page dimensions

Workflow

0. Identify Scenario (new)

Before any other step, identify the presentation scenario. Read references/content-scaffolding.md for full templates.

Each template defines: slide count, page titles, required content per page.

0.5 Bilingual Layout (if needed)

If the audience is bilingual (e.g. HK, Singapore, global Chinese conference), or the user mentions Chinese + English:

• Use the bilingual layout mode from references/content-scaffolding.md
• Heading: English (large) + Chinese subtitle (smaller, --text-muted)
• Body bullets: Chinese first, English parenthetical optional
• Avoid pure English-only decks for HK/TW/SG audiences

1. Plan the Deck

Define slide count and content per slide based on the scenario template. Each slide = one <section class="slide">.

1.5 Art Direction (run this before building)

Run this step whenever the user hasn't provided a specific visual style. Read skills/slide-creator/references/art-direction.md for the full style taxonomy, CSS token templates, and style-brief output format.

Step A — Ask 4 questions:

1. Audience & setting: who is this for, and in what context (investors / internal team / public keynote)?
2. Mood keywords: how should the audience feel (authoritative / energetic / friendly / geeky-modern)?
3. Brand constraints: any required brand colors, logo, or fonts?
4. Reference material: any template references to align with (images, web links, existing deck screenshots)?

Step B — Generate a visual style picker page: Do NOT present style options as text descriptions — users can't evaluate styles from words alone.

Reference handling rules:

• If user provides image files/screenshots: sample palette (primary/surface/accent), inspect layout density, corner radius, typography tone from the visual.
• If user provides web URLs: use web_fetch to extract design cues. Critical — follow this extraction protocol to avoid misreading the style:
  1. Ignore the brand name / domain name — never infer visual style from the product's industry or name (e.g. "Neo" does NOT mean neon, "Opera" does NOT mean European luxury).
  2. Read copy tone & vocabulary — the words used on the page reveal mood (e.g. "surgical precision", "quiet confidence" → restrained; "unleash", "radically" → bold/aggressive).
  3. Extract explicit color vocabulary — look for CSS keywords in the fetched text, or color names mentioned in body text / alt tags. Warm vs cool, light vs dark, muted vs saturated.
  4. Infer layout density — count words per section; sparse = editorial/luxury, dense = technical/functional.
  5. Identify decorative motifs — mentioned or implied (e.g. geometry, gradients, photography, illustration, line art, brutalism).
  6. Cross-check against art-direction.md — find the closest matching template, then describe the delta (e.g. "Style G but warmer, replace blue with burnt orange, add subtle grid lines").
  7. When uncertain: be conservative — under-promise the style match and present 3 options where Option A is your best interpretation, B is safer/cleaner, C is more experimental. Never confidently assert a style that contradicts the actual visual evidence.
• If user provides both: prioritize image cues first, URL cues second.
• If no references are provided: use built-in style taxonomy defaults.

Then:

1. Create output/style-picker/index.html — a single page with 3 side-by-side mini slide previews (16:9 aspect ratio), each fully rendered with real CSS (colors, fonts, layout, decorative elements). Each preview must look like an actual slide, not a color chip.
2. Build these 3 options as: (A) Reference-faithful, (B) Safer corporate variant, (C) Bolder creative variant.
3. preview(action='serve') the directory and show the preview URL.
4. Each card has a label below: style name + one-line description.
5. Add onclick highlight so the user can click to indicate their choice.

User picks by saying "Choose A" / "I want B" / "Blend A+C" etc.

Step C — Generate style-brief.md: Once user selects a style, write a style-brief.md (template in art-direction.md) in the project directory. All subsequent HTML/CSS work must follow this brief.

Step D — Ask for brand assets (logo / colors): After user picks a style, ask:

"Do you have a logo or brand color to include? You can upload an image file, and I’ll embed the logo across the slides."

If logo uploaded: embed as base64 in HTML (use base64.b64encode in bash), place in top-left or top-right corner at ≤60px height. If brand color given: override --accent in CSS token block with user's color.

2. Choose a Theme

If Art Direction was completed, the style-brief.md is the theme spec — skip this table. Otherwise, use as a quick fallback:

3. Build HTML + CSS

Create a project directory with index.html + styles.css.

Start from assets/base.css — structural skeleton (slide dimensions, print rules, layout helpers) with NO colors or fonts. Layer your theme on top:

/* Example theme layer — customize freely */
body {
  font-family: 'Inter', sans-serif;
  color: #fff;
  background: #000;
}
.slide { background: #0a0a0a; }
.slide-tag { background: rgba(0,120,255,0.15); color: #0078ff; }
.card { background: rgba(255,255,255,0.04); border: 1px solid rgba(255,255,255,0.08); }

Mandatory structural rules (in base.css — don't remove):

.slide { width: 1280px; height: 720px; page-break-after: always; overflow: hidden; }
@page { size: 1280px 720px; margin: 0; }

Key rules:

• Use px units — never vh/vw/rem/% for slide dimensions
• Google Fonts: use <link> in <head>, export script waits for network idle
• Viewport meta: <meta name="viewport" content="width=1280">
• Content must fit within 720px height — overflow is clipped

4. Preview (Optional)

Use preview(action='serve') to preview in browser before exporting.

5. Export to PDF

python3 skills/slide-creator/scripts/export_pdf.py --dir <project-dir> --output output/<name>.pdf

Options:

• --dir — directory containing index.html (required)
• --output / -o — output PDF path (default: <dir>/deck.pdf)
• --width — slide width in px (default: 1280)
• --height — slide height in px (default: 720)

6. Verify

The script prints slide count and confirms output path. Extra check:

import fitz
doc = fitz.open("output/deck.pdf")
print(f"Pages: {doc.page_count}")
for p in doc:
    r = p.rect
    print(f"  {r.width*96/72:.0f}x{r.height*96/72:.0f}px")

Style Guidelines

• No default brand — every deck gets a theme tailored to its content
• Prefer Art Direction-first workflow (questions → references → user pick → style-brief.md)
• Ask the user for preference: dark/light, accent color, font, mood
• Each slide should have clear visual hierarchy: tag → title → content
• Keep text concise — slides are visual, not documents
• Use .bg-glow with theme-colored radial gradients for depth
• Do not use web_search for style exploration by default; prioritize user-provided reference images/links plus templates in art-direction.md.
• When the user provides reference links, use web_fetch to extract design cues (color tone / voice / layout), but implement final CSS using local templates and token variables.
• When the user asks for "art-direction advice / style advice", always generate a visual style-picker preview page (3 options) — never rely on text-only descriptions.
• Build HTML strictly against chosen style brief, then export PDF (do not skip brief unless user explicitly opts out)
• After style is chosen, always ask about logo / brand assets before building
• For HK/TW/SG/bilingual audiences, default to bilingual layout unless user says English only

Style Microtweaks (after style is chosen)

If user says "change primary color to red" / "switch the font" / "increase corner radius", do NOT restart art direction. Instead, directly patch the --accent / --font-head / --radius CSS variable in styles.css. Only restart art direction if user wants a completely different style.

Gotchas

• Chromium: pre-installed via workspace/setup.sh on container start (~641 MB cached at ~/.cache/ms-playwright/). Do NOT run playwright install on every export — it re-downloads the same browser. Only run it if export_pdf.py fails with Executable doesn't exist, and in that case also append the command to workspace/setup.sh so it persists.

• Fonts: Google Fonts need HTTP — the export script starts a local server automatically

• Emoji rendering: headless Chromium may lack emoji fonts — use SVG icons instead

• Large images: embed as base64 or use relative paths (local server serves the project dir)

• Slide overflow: content exceeding 720px height is clipped — design within bounds

• ⚠️ PDF text not selectable (verified 2026): filter / backdrop-filter CSS on any ancestor containing text causes Chromium to rasterize that layer to bitmap during PDF export — all child text becomes pixels, not selectable. Fix: NEVER apply filter/backdrop-filter to containers holding text. Only apply to empty decorative <div> elements (e.g. .blur-layer, .glow-overlay) with no text children. Same rule applies to mix-blend-mode on text parents. Pre-export checklist: search HTML for filter/backdrop-filter on non-decorative elements and strip them.

• ⚠️ Footer / source attribution — use standard component + strict bottom-safe-area contract (verified 2026): ad-hoc footer markup causes inconsistent positioning across slides. Always use this .slide-footer pattern for all source citations, page numbers, and disclaimers.

  Hard layout contract (do not skip):

  • Every non-cover slide must have a dedicated content wrapper (e.g. .slide-body) that reserves footer space.
  • Hard rule: .slide-body must reserve footer space with bottom-safe-area >= 96px (default 96px). Example: .slide-body { padding: 52px 72px 96px; }.
  • Footer must be outside normal flow: position: absolute; bottom: 24px, as a sibling of .slide-body directly under .slide.
  • Never place source/disclaimer text inside .slide-body.
  • Cover page can be exception only when it has no .slide-footer.

  This enforces physical separation: body content area ends above a reserved footer lane, and footer stays in the canvas-bottom lane, so they never overlap regardless of content density.

  <footer class="slide-footer">
    <span class="footer-source">Source: CoinGecko · Coinglass · DefiLlama</span>
    <span class="footer-page">03 / 12</span>
  </footer>
  

  .slide-footer {
    position: absolute;
    bottom: 24px; left: 48px; right: 48px;
    display: flex; justify-content: space-between; align-items: center;
    font-size: 11px; color: rgba(255,255,255,0.35);
    border-top: 1px solid rgba(255,255,255,0.08);
    padding-top: 8px;
    z-index: 2;
  }
  .slide > *:not(.bg-glow):not(.slide-footer) {
    position: relative;
    z-index: 1;
  }
  .slide-body { padding-bottom: 96px; }
  

  Never use in-flow / position: relative footers — they shift when slide content height changes. Also exclude .slide-footer from generic .slide > * stacking rules, or CSS order can accidentally override footer layering. Validation checklist (required before export):

  1. every non-cover slide has .slide-body and .slide-footer as siblings;
  2. .slide-body bottom padding is >=96px;
  3. source/disclaimer text appears only inside .slide-footer, never in body containers.

• ⚠️ z-index / decorative overlay bug (verified 2026): .bg-glow and other decorative pseudo-layers MUST be positioned with z-index: 0 and all real slide content given z-index: 1 explicitly. If .bg-glow is a sibling of .slide > * (not a ::before/::after pseudo-element), add this rule to ensure content is never visually buried:

  .slide > *:not(.bg-glow) { position: relative; z-index: 1; }
  .bg-glow { position: absolute; z-index: 0; pointer-events: none; }
  

  Failure mode: PDF exports show content pushed to the bottom or invisible, even though browser preview looks fine (browser compositing handles z-order more forgivingly than Chromium's print path).