High-Quality Slides Skill

Goal: replicate Genspark AI Slides' "feels designed, not templated" quality bar inside Claude. Optimized for the decisive content + curated visual philosophy, not the "wall of bullet points on a stock theme" pattern.

This skill is opinionated. Do not skip phases. Most low-quality AI decks fail at Phase 1–3 (no research, no narrative, no visual concept) and overinvest in Phase 4 (typography polish on garbage content). This skill inverts that.

Operating principle

A great deck = research × narrative × visual system × per-slide layout decisions. If any one factor is zero, the deck is zero. The model's job is to be a slide art director, not a content shoveler.

Three rules that override everything else:

1. One idea per slide. If a slide has more than one assertion, split it.
2. Visual relationship > text list. Whenever bullet points appear, ask "can this be a 2×2, timeline, comparison table, stat strip, or diagram?" Default answer: yes.
3. Pixel-perfect, viewport-locked. Every slide fits in a fixed canvas (1920×1080 default). No overflow, no scroll-within-slide, no responsive text reflow.

Five-phase workflow (mandatory order)

Modeled on Genspark's Guide Mode. Each phase has a gate — do not advance until the gate passes.

Phase 1 — STRATEGY (audience & purpose)

Ask, or infer from context, then state back in one paragraph:

• Who is the audience? (board / customer / team / conference / academic)
• What is the single decision or action the audience should take after seeing this?
• What's the room? (live presentation, sent as PDF, embedded on web)
• Time budget for delivery? (drives slide count)
• Brand constraints? (logo, color, font, existing template)

Gate: if audience or call-to-action is unknown, ask the user with AskUserQuestion. Do not proceed.

Phase 2 — SUBSTANCE (research & evidence)

Collect the raw material before designing anything. Never invent statistics or quotes.

Sources, in order of priority:

1. Files the user uploaded — read every one (Read tool, OCR via Python if needed).
2. URLs the user provided — fetch.
3. Connected MCP data (Slack, Asana, Drive, analytics).
4. WebSearch — only for facts the user couldn't supply, and only with citations.

Produce a research-notes.md scratchpad in the outputs folder with: claims, numbers, quotes, source URLs, image candidates (URLs or descriptions). This file is throwaway but you must produce it — it forces evidence discipline.

Gate: every assertion that will appear on a slide has a source in research-notes.md.

Phase 3 — STRUCTURE (narrative & outline)

Choose a narrative arc that fits the goal. Default arcs:

Output a numbered outline. For each slide, write three things only:

1. Headline assertion (≤ 12 words, declarative — "Revenue doubled in Q3", NOT "Q3 Revenue")
2. The one supporting fact (a number, a quote, a comparison)
3. The intended visual (chart type / diagram / photo / icon set / typographic)

Gate: user reviews the outline. Do NOT design slides before the headline list is approved. Use AskUserQuestion to confirm or ask for edits.

Phase 4 — DESIGN SYSTEM (style & density)

Choose the visual system once, apply consistently. Do not "design each slide" — design the system, then instantiate slides from it.

Required design tokens (write to design-tokens.css as CSS custom properties):

:root {
  /* Canvas */
  --slide-w: 1920px;
  --slide-h: 1080px;
  --safe-pad: 96px;        /* outer padding, ~5% of width */

  /* Type scale — pick ONE display font + ONE text font */
  --font-display: "Inter", system-ui, sans-serif;
  --font-text:    "Inter", system-ui, sans-serif;
  --fs-hero:      120px;   /* title slide */
  --fs-h1:        72px;    /* slide headline */
  --fs-h2:        44px;
  --fs-body:      28px;
  --fs-caption:   20px;
  --lh-tight:     1.1;
  --lh-body:      1.4;

  /* Color — pick ONE base + ONE accent. No more. */
  --bg:        #0B0B0C;
  --fg:        #F5F4F0;
  --muted:     #8C8A85;
  --accent:    #FF5A1F;
  --surface:   #17171A;
  --line:      rgba(255,255,255,.12);

  /* Grid */
  --col: 12;
  --gutter: 32px;
}

Pick a style preset from layouts.md (see attached). Presets encode color + type + grid + motion, e.g. bold-signal, swiss-modern, editorial-serif, dark-botanical, neon-cyber, notebook-tabs. If the user can't articulate a preference, generate 3 preset previews (single hero slide for each) and let them pick — this is the "show, don't tell" trick that makes the user feel ownership of the design.

Density rules (override per-slide judgement only with reason):

• Title ≤ 12 words. Body ≤ 30 words OR ≤ 5 list items.
• One hero element per slide (chart OR quote OR photo OR diagram — not all four).
• Whitespace ≥ 30% of canvas.

Gate: preset chosen, tokens written, single sample slide approved before mass generation.

Phase 5 — BUILD (generate slides)

Generate slides as HTML, one <section class="slide"> per page, single self-contained file, inline CSS/JS, viewport-locked. Use the patterns in html-template.md.

For each slide in the outline:

1. Choose a layout pattern from layouts.md based on the intended visual: cover, headline-only, three-stat-strip, quote-hero, two-column-compare, timeline, process-flow, 2x2-matrix, chart-focus, image-with-caption, data-table-clean, kpi-grid, section-divider, closing-cta.
2. Drop in the headline + supporting fact + visual.
3. Render the visual:

• Charts → inline SVG (preferred) or Chart.js / D3 from CDN. Real data only. Strip chartjunk: no gridlines unless necessary, no 3D, no default rainbow palettes — use the accent color + neutrals.
• Diagrams → hand-built SVG or HTML/CSS boxes with the design tokens. Avoid mermaid for hero diagrams (looks generic); use it only for technical flow appendices.
• Icons → Lucide via CDN, OR inline SVG. Consistent stroke width across the deck.
• Photos → only if the user provided them, or via clearly-licensed source. Apply a duotone or grayscale-with-accent treatment so stock images don't break the system. If no real photo, prefer a typographic/abstract treatment over a generic AI image.
• Smart visualization (Genspark's edge): match content semantics to visual form. Code → terminal mockup. Comparison → side-by-side cards. Process → arrowed steps. Numbers → big number + tiny label. Quote → oversized punctuation. This is the single biggest quality lever.

Then run the self-check before delivery:

• [ ] Every slide has exactly one headline assertion (declarative, ≤ 12 words).
• [ ] No two slides use the exact same layout in a row (unless intentional sequence).
• [ ] All numbers/quotes trace to research-notes.md.
• [ ] No text overflows the safe area at 1920×1080.
• [ ] Color usage: ≤ 1 accent color, used sparingly (< 20% of pixels).
• [ ] No emoji unless the brief is informal.
• [ ] Final slide has a clear CTA / next step (matches Phase 1 audience action).

If a check fails, fix the slide, do not ship.

Optional: pptx export

If the user needs an editable .pptx:

1. First produce the HTML deck (it's the design source of truth).
2. Then call the pptx skill, passing the outline + design tokens + per-slide layout choices as the construction plan. The HTML serves as the visual reference.
3. Note in the handoff: complex SVG diagrams may need to be rasterized for pptx compatibility — call this out to the user.

Anti-patterns (these are what makes AI decks look like AI decks)

Refuse these by default. If the user insists, comply but note the trade-off.

• Purple-to-blue gradients on white. Generic.
• Stock photos of diverse people pointing at laptops.
• 7+ bullet points per slide.
• "Title Case Every Word" headlines that don't make an assertion.
• Decorative AI-generated images that don't carry information.
• Same icon pack used for unrelated concepts.
• Default Chart.js color palette.
• Footer with date + page number + company name + tagline (pick one).
• "Thank You / Questions?" closing slide — replace with a real CTA.

Files in this skill

Load files on demand. Do not preload everything into context.

Quick reference — what the agent actually does, step by step

1. Parse the brief → identify audience, action, format, slide count, brand.
2. If anything in (1) is unknown → AskUserQuestion. Do not start.
3. Read all attached files + fetch URLs → write research-notes.md with sources.
4. Propose narrative arc + numbered outline (headline + fact + visual per slide).
   → AskUserQuestion to confirm outline. Iterate until approved.
5. Propose 2–3 style preset previews → user picks one.
6. Write design-tokens.css. Build sample slide. Confirm with user.
7. Generate full HTML deck, one layout-pattern per slide based on visual intent.
8. Run the self-check list. Fix anything that fails.
9. Save deck.html + research-notes.md + design-tokens.css to outputs folder.
10. (Optional) Export to .pptx via the pptx skill, using the HTML as the spec.
11. Hand back computer:// links. Offer one round of revisions.

Last note: the user does not care about your process. They care about getting a deck that looks like a designer made it. The phases exist to force that outcome — keep them, but communicate progress in plain language, not jargon.