A Claude Code skill that produces decks worth sending — not "AI-looking" slide soup.
Inspired by reverse-engineering Genspark AI Slides and similar high-quality AI presentation tools, this skill teaches Claude to behave like a slide art director: research first, narrative second, system third, slides last.
The goal is simple: when someone says "make me a deck," the output should look designed, not generated.
When triggered, Claude follows a 5-phase workflow with mandatory gates between phases. The phases are deliberately frontloaded — most low-quality AI decks fail by skipping straight to "make pretty slides" without doing research or narrative work first.
| Phase | Output | Gate |
|---|---|---|
| 1. Strategy | Audience, action, format, slide count | Asks the user if anything is unclear |
| 2. Substance | research-notes.md — every fact with a source |
No invented stats |
| 3. Structure | Numbered outline: headline + supporting fact + visual per slide | User approves outline |
| 4. Design system | design-tokens.css, style preset choice, sample slide |
User approves sample |
| 5. Build | deck.html, self-contained, viewport-locked, keyboard-navigable |
Self-check passes |
The deliverable is an HTML deck by default. Optionally exports to .pptx by composing with Claude's pptx skill.
- Pixel-perfect rendering, no PowerPoint version drift
- Keyboard navigation + deep linking via URL hash
- Self-contained single file — no build step, no dependencies
- Trivial to convert to PDF or screenshots for sharing
- HTML/CSS is what Claude is best at writing
If you need editable .pptx for a colleague who lives in PowerPoint, the skill can hand off to the pptx skill using the HTML as the design spec.
git clone https://github.com/andyqiu847-ai/high-quality-slides.git ~/.claude/skills/high-quality-slidesRestart your Claude Code session. The skill will be auto-discovered.
cd your-project
mkdir -p .claude/skills
git clone https://github.com/andyqiu847-ai/high-quality-slides.git .claude/skills/high-quality-slides
git add .claude/skills/high-quality-slides
git commit -m "Add high-quality-slides skill"Everyone on the team gets the skill on next git pull.
ls ~/.claude/skills/high-quality-slides/
# SKILL.md layouts.md html-template.md README.md LICENSEIn Claude Code, open a new session and ask: "What skills do you have available for making slides?" — it should list this one.
Just describe what you want naturally. The skill auto-triggers on prompts like:
- "做一份关于 X 的高质量 PPT"
- "Make me a Genspark-style pitch deck for our Q4 review"
- "Turn this report into a 10-slide presentation" (with a PDF/docx attached)
- "I need an investor deck for tomorrow"
Claude will then walk you through Phase 1 questions (audience, action, length), do its research, propose a narrative arc, ask you to pick a visual style from a few previews, and only then start building slides.
You can also short-circuit if you know what you want:
"Make a 12-slide HTML deck in the
swiss-modernpreset about our user research findings. Audience is the design team. Sources are in~/Documents/research-notes/."
high-quality-slides/
├── SKILL.md # Workflow + gates + anti-patterns (loaded on trigger)
├── layouts.md # 14 layout patterns + 12 style presets (Phase 4–5)
├── html-template.md # HTML scaffold, chart rendering, export paths (Phase 5)
├── README.md # This file
└── LICENSE
The skill uses progressive disclosure — SKILL.md is a ~180-line map, and the larger reference files load only when their phase begins. This keeps Claude's context window lean and the routing predictable.
cover, section-divider, 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, closing-cta.
Each pattern is a recipe with a clear "when to use this" rule. Claude picks per slide based on the content's semantic shape — comparisons get two-column-compare, processes get process-flow, single decisive numbers get headline-only.
Dark: bold-signal, electric-studio, creative-voltage, dark-botanical.
Light: notebook-tabs, pastel-geometry, split-pastel, vintage-editorial.
Specialty: neon-cyber, terminal-green, swiss-modern, paper-and-ink.
Each preset encodes color palette, type pairing, grid system, and motion. They're real design systems, not vague style labels.
Three rules that override everything else:
- One idea per slide. If a slide has more than one assertion, split it.
- Visual relationship > text list. Whenever bullets show up, ask: can this be a 2×2, timeline, comparison, stat strip, diagram? Default answer: yes.
- Pixel-perfect, viewport-locked. Every slide fits in 1920×1080. No overflow, no internal scroll.
Plus an opinionated list of anti-patterns (see SKILL.md), including the canonical AI-slop signals:
- Purple-to-blue gradients on white
- Stock photos of diverse people pointing at laptops
- 7+ bullet points per slide
- "Title Case Every Word" headlines that don't assert anything
- Default Chart.js color palettes
- "Thank You / Questions?" closing slides
This skill cooperates with other Claude skills:
pptx— for editable PowerPoint export. The HTML deck serves as the design spec.pdf— for reading source documents in Phase 2.xlsx— for pulling numbers out of spreadsheets in Phase 2.docx— for ingesting Word reports.
A parent agent in the Claude Agent SDK can also spawn a subagent with this skill enabled to delegate "make a deck" as a unit of work and get back a deck path.
Pull requests welcome on any of these:
scripts/generate_deck.py— programmatic entry point for non-chat usageexamples/— a few decks produced with public-topic prompts, to serve as visual references- More layout patterns: split-screen video stills, before/after with sliders, animated chart reveal
- More presets: brand-aware preset that ingests a logo + 2 colors and adapts the system
- MCP server wrapper, so non-Claude agents can call it as a tool
The progressive-disclosure architecture and "show, don't tell" style discovery idea are borrowed from frontend-slides — read that repo, it's excellent.
The 5-phase Guide Mode workflow is reverse-engineered from Genspark AI Slides Guide Mode.
The smart-visualization rules (content semantics → visual form) come from observing how Genspark, Gamma, and human designers actually make per-slide layout decisions on real decks.
MIT. See LICENSE.