Slide Creator

Slide Creator

Generate zero-dependency HTML presentations that run entirely in the browser. This skill guides users from raw ideas to a polished, animated slide deck — using visual style discovery ("show, don't tell") to nail the aesthetic before writing a line of code.

Core Philosophy

1. Zero Dependencies — Single HTML files with inline CSS/JS. No npm, no build tools.
2. Show, Don't Tell — People can't articulate design preferences until they see options. Generate visual previews rather than asking abstract questions like "do you want minimalist or bold?"
3. Distinctive Design — Avoid generic AI aesthetics (Inter font, purple gradients, predictable heroes). Every deck should feel custom-crafted.
4. Viewport Fitting — Slides must fit exactly in the viewport with no scrolling. Content that overflows gets split across slides, not squished. A presentation that scrolls mid-slide is broken — this is why we enforce density limits.
5. Plan Before Generate — For complex decks, --plan creates a PLANNING.md outline first; --generate then produces HTML. Separating thinking from execution leads to better structure and less backtracking.

Command Flags

Parse the invocation to determine mode:

• --plan [prompt] — Planning mode. Inspect resources/, analyze the prompt, create PLANNING.md. Stop — do NOT generate HTML.
• --generate [instructions] — Generation mode. Read PLANNING.md if present (skip Phase 1/2 questions), then generate HTML.
• --export pptx [--scale N] — Export the most recently modified HTML as PPTX via the bundled Python script. Requires Python 3 + pip install playwright python-pptx. No Node.js required.
• No flag — Auto-detect mode (Phase 0).

---

Planning Mode (--plan)

1. Scan resources/ — read text/markdown files, note images. Tell the user what was found (or "Planning from prompt only" if empty).
2. Extract: topic, audience, tone, language, slide count, goals from the prompt.
3. Draft the plan following references/planning-template.md.
4. Save as PLANNING.md in the working directory.
5. Present slide count, structure, and key decisions. Ask for approval.
6. Stop. Do NOT generate HTML.

---

Export Mode (--export pptx)

1. Find *.html in current directory (prefer most recently modified).
2. Run: python3 <skill-path>/scripts/export-pptx.py <presentation.html> [output.pptx]
3. Report the PPTX file path and slide count.

The script uses Playwright with the user's existing system Chrome (--channel chrome flag, no Chromium download). It captures pixel-perfect screenshots of each slide and assembles them into a PPTX. Only pip install playwright python-pptx required — no Node.js, no 300MB browser download. If system Chrome is not found, the script will report an error and ask the user to install Chrome.

---

Phase 0: Detect Mode

Determine what the user wants:

• Mode A — New Presentation: Check for PLANNING.md first. If it exists, read it as source of truth and jump directly to Phase 3 — skip Phase 1/2.
• Mode B — PPT Conversion: User has a .ppt/.pptx file → go to Phase 4.
• Mode C — Enhance Existing: Read the existing HTML, understand its structure, then enhance. When adding content, always check viewport fit — if adding would overflow a slide, split the content rather than cramming it in. Proactively split and inform the user.

Content-type → Style hints (use when user hasn't chosen a style):

---

Phase 1: Content Discovery

First, silently scan for a resources/ folder. If found, read text/markdown files and note images as background context. Don't ask the user to take any action.

Then gather everything in a single AskUserQuestion call with all 5 questions at once — collecting everything before submitting avoids back-and-forth.

• Purpose (single select): Pitch deck / Teaching+Tutorial / Conference talk / Internal presentation
• Length (single select): Short 5-10 / Medium 10-20 / Long 20+
• Content (single select): All content ready / Rough notes / Topic only
• Images (single select): No images / ./assets / Other (user types path)
• Inline Editing (single select): Yes — edit text in-browser, auto-save (Recommended) / No — presentation only

Default: Always include inline editing unless user explicitly selects "No". When --generate skips Phase 1, include inline editing by default.

If user has content, ask them to share it after submitting the form.

Language: Detect from the user's message — never default to a fixed language. Maintain the detected language throughout all slide text including labels, CTAs, and captions. English may appear as secondary annotation text only when the style calls for it.

Image Evaluation

Skip if user chose "No images." Text-only decks are fully first-class — CSS-generated gradients, shapes, and typography create compelling visuals without any images.

If images are provided:

1. ls the folder, then use Read (multimodal) to view each image.
2. For each image: mark USABLE or NOT USABLE (with reason: blurry, irrelevant, broken) + what it represents + dominant colors + shape.
3. Build a slide outline that co-designs text and images from the start. This is not "plan slides, then fit images in after." Example: 3 usable product screenshots → 3 feature slides anchored by those screenshots.
4. Present the evaluation and proposed outline, then confirm via AskUserQuestion (Looks good → Style B / Adjust images / Adjust outline).

---

Phase 2: Style Discovery

Most people can't articulate design preferences in words. Generate 3 mini visual previews and let them react — this is the "wow moment" of the skill.

Style Path

Ask via AskUserQuestion:

• "Show me options" → ask mood question → generate 3 previews based on answer
• "I know what I want" → show preset picker (Bold Signal / Blue Sky / Modern Newspaper / Neo-Retro Dev Deck — with "Other" option for all 21 presets)

Before showing presets, silently scan <skill-path>/themes/ for subdirectories. Skip any directory whose name starts with _ (those are examples/templates, not real themes). Each remaining subdirectory with a reference.md is a custom theme. Append them to the preset list as Custom: <folder-name> entries. If any custom themes exist, mention them first: "I also found N custom theme(s) in your themes/ folder."

Available Presets (full details in STYLE-DESC.md):

Mood → Preset mapping:

Generate Previews

Create 3 mini HTML files in .claude-design/slide-previews/ (style-a/b/c.html). Each is a single title slide (~50-100 lines, self-contained) demonstrating typography, color palette, and animation style.

If a USABLE logo was found in Step 1.2, embed it (base64) into each preview — seeing their own brand in 3 different aesthetics makes the choice feel personal, not abstract.

Never use: Inter/Roboto/Arial as display fonts, generic purple-on-white gradients, predictable centered hero layouts.

Present the 3 files with a one-sentence description each, then ask via AskUserQuestion which they prefer (Style A / B / C / Mix elements).

---

Phase 3: Generate Presentation

Generate the presentation based on content from Phase 1 and style from Phase 2. If PLANNING.md exists, it's the source of truth — skip Phases 1 and 2.

If Blue Sky style is selected → use the starter template

Read references/blue-sky-starter.html and use it as the base.

All 10 signature visual elements (orbs, noise texture, grid, glassmorphism, spring-physics horizontal track, gradient text, cloud hero, pill dots, counter, entrance animation) are already implemented in that file. You only need to:

1. Set --slide-count in :root to your actual slide count
2. Replace the example slides in #track with your content, following the slide-type patterns in the comment block
3. Fill the orbPositions array in JS — one [l1,t1,s1, l2,t2,s2, l3,t3,s3] entry per slide

Do not rewrite the visual system CSS. Do not change the track/slide layout. Content goes inside .slide wrappers using the pre-built component classes: .g (glass card), .gt (gradient text), .pill, .stat, .divider, .cols2/3/4, .ctable, .co (amber callout), .warn (red warning), .info (blue info), .layer (org row), ul.bl (bullet list).