Interactive Deck Builder

A framework for self-contained HTML presentations: one canvas, many Three.js scenes, click-driven interactions per slide, modal CLI / REPL playback simulator, dark accent-colored design system, live-data tiles fed from any source. Topic-agnostic — bring your own metaphor, your own data, your own scenes.

When to use

• The user asks for an interactive HTML / browser presentation, deck with games, or "talk with 3D animations" on any topic.
• The user wants Three.js scenes to illustrate a concept (any concept — process flow, market dynamics, scientific phenomenon, system architecture, narrative metaphor).
• The user wants slides that show real data baked in from an external source (CSV file, REST API, SQL query, kubectl output, file system, spreadsheet, anything).
• The user already has a deck built with this framework and wants a new scene / section / animation added.

Skip this skill when:

• The user wants a static deck (Keynote, PowerPoint, Google Slides, plain Markdown). For those, use a presentation-design / animation-principles skill.
• The user wants a tech-talk outline without the framework — use a presentation-outline skill.
• The user wants pure Three.js fundamentals (scene graph, materials, cameras) without the deck wrapper — use a threejs-webgl skill.

This skill is the framework. Bring complementary skills for content writing, animation theory, design, and Three.js fundamentals.

Required structure

Every deck is a self-contained directory served as static files:

<deck-name>/
├── index.html        # <section class="slide" data-section="X" data-color="Y" data-scene="Z"> per slide
├── styles.css        # design system: --accent-* colors, .slide, .scene-slide, .fullscreen-scene
├── presentation.js   # slide controller, keyboard nav, build steps, scene wiring
├── scenes.js         # Three.js scenes — one per data-scene name, registered in `scenes` map
├── demo.js           # CLI / REPL / terminal playback simulator (modal, optional)
├── data/             # CSV / JSON / etc. shown on slides AND fed into scenes (optional)
├── code-examples/    # snippets shown on slides AND used by the demo simulator (optional)
├── README.md         # how to open + section map
├── OUTLINE.md        # section-by-section structural map
└── SPEAKER-GUIDE.md  # per-slide talking points (auto-generated from <aside class="notes">)

Run with python3 -m http.server 8765 from the parent directory, then open http://localhost:8765/<deck-name>/.

Workflow

1. Pick the topic and narrative beats. Map your story to ~6-9 sections. Each section gets one accent color (teal / red / purple / amber / green / blue / cyan / rose / deep-purple). The colors are mood markers, not topic markers — pick what fits the beat (red = problem / alarm, green = fundamentals / safety, blue = scale / depth, cyan = open / cool, rose = warmth / data, etc.).
2. Capture real data first (optional but high-value). Whatever your topic, find the real numbers that ground it. Sales deck → real revenue / churn / LTV. Lecture → real survey / experiment results. System talk → real cluster output. Product demo → real usage stats. Load references/data-sources.md for the discovery patterns by data source type.
3. Copy framework files from assets/templates/ into the new deck folder: index.html, styles.css, presentation.js, scenes.js, demo.js. Or run bash scripts/new-deck.sh <deck-name> to scaffold.
4. Author slides in index.html as <section class="slide" data-section="..." data-color="..." [data-scene="..."]>. Use the established slide types: title, big-statement, framework, code, divider, scene, recap, q-and-a. Load references/slide-types.md for the full template gallery.
5. Build Three.js scenes for concepts that benefit from interaction. Each scene is a function in scenes.js that returns { scene, camera, update, ... } and is registered in the scenes map. The shipped templates ship ONLY two ambient scenes — particles (title) and aurora (dividers) — so every interactive scene is built per deck, not copied from a shipped library. Load references/scene-recipes.md for structural recipes covering ~12 common patterns (race, particle-flow, click-to-destroy, density / fill-up, slider scaling, force-directed graph, wave generator, topology, color-shift, hub-and-spoke, migration, comparison split-screen). Load references/threejs-scenes.md for the factory contract + accent-color sync + raycasting + scene-local-vs-world-coords gotcha. For Three.js fundamentals (geometries, materials, lighting, performance), defer to a threejs-webgl skill.
6. Wire each scene to its slide — data-scene="<name>" on the section + matching control elements + a refreshXxxUI() function in presentation.js fed by a callback the scene exposes via setOnChange(). Load references/scene-wiring.md for the wiring contract.
7. Add demo simulator entries in demo.js (optional) — one per scripted walkthrough. Each entry is an array of { say } (narration) and { run, out, status } (commands + simulated output). Triggered with the D key. Generic enough for any CLI: kubectl, psql, redis-cli, git, npm, AWS CLI, custom REPLs, or scripted shell.
8. Generate the speaker guide with python3 scripts/extract-speaker-notes.py <deck-dir> after every slide-content change.
9. Add the fullscreen-scene toggle to every scene slide. Load references/fullscreen-scene-pattern.md if a slide's right panel overflows the viewport in fullscreen mode — the fix is position: fixed + grid minmax(0, ...) + overflow: hidden.
10. Validate by opening in a browser. Don't trust the build alone. Walk every slide with arrows, toggle fullscreen on every scene, open the demo modal on D, confirm section colors render distinctly. Load references/qa-checklist.md for the full pre-talk list.

Available resources

• assets/templates/index.html — slide skeleton with persistent UI frame, help overlay, demo overlay, notes panel, overview grid.
• assets/templates/styles.css — full design system: --accent-* colors, .slide, .scene-slide, .fullscreen-scene overrides.
• assets/templates/presentation.js — slide controller, keyboard nav, build-step framework, scene wiring.
• assets/templates/scenes.js — slim Three.js scene framework: shared renderer + animation loop + scene registry + two ambient scenes (particles, aurora). All other scenes are built per deck.
• assets/templates/demo.js — modal CLI / REPL / terminal playback simulator.
• assets/examples/README.md — pointers to example decks and how the patterns transfer to non-technical topics.
• scripts/extract-speaker-notes.py — generates SPEAKER-GUIDE.md from <aside class="notes"> blocks.
• scripts/new-deck.sh — scaffold a new deck directory by copying framework files and substituting the deck name.
• scripts/discover-cluster.sh — preset data discovery for kubectl-style sources (one example among many; see references/data-sources.md for patterns covering CSV / JSON / API / SQL / file-scan).
• references/design-system.md — colors, typography, layout, accent-color rules.
• references/slide-types.md — title, big-statement, framework, code, divider, scene, recap, q-and-a templates.
• references/threejs-scenes.md — scene factory contract, render loop, accent-color sync, raycasting, scene-local-vs-world-coords gotcha, teardown disposal. Defers Three.js fundamentals to a threejs-webgl skill; defers per-pattern recipes to scene-recipes.md.
• references/scene-recipes.md — structural recipes for ~12 common scene patterns (race, particle-flow, click-to-destroy, density, slider-scaling, force-graph, wave, topology, color-shift, hub-and-spoke, migration, comparison). Each recipe gives shape, gotchas, implementation sketch, UI tile pattern.
• references/scene-wiring.md — data-scene, refreshXxxUI, build steps, the integration contract.
• references/data-sources.md — patterns for CSV / JSON / API / SQL / kubectl / file-scanning / spreadsheets — what to bake in, how to bake it in, freshness strategies.
• references/fullscreen-scene-pattern.md — toggle button + CSS overrides for the bottom-bar layout.
• references/qa-checklist.md — pre-talk verification list (topic-agnostic).

Top gotchas (always inline — do not skip)

• Three.js scene-local vs world coords. When a scene is offset (e.g. scene.position.y = 3.5 to push it into the upper viewport), child positions are scene-local. If you target an object via obj.getWorldPosition(wp), you must scene.worldToLocal(wp) before assigning to a particle whose movement is computed in scene-local space. Otherwise particles aim N units off and never visually arrive.
• Fullscreen-scene CSS must use position: fixed + grid minmax(0, ...). position: absolute + flex inside the slide can overflow the viewport on the right. Use min-width: 0; overflow: hidden on grid children to prevent text from blowing out the bar.
• canvas.style.pointerEvents defaults to none so clicks pass through to slides. Toggle to 'auto' only on slides that need raycasting. Forgetting this means the user can't click anything in the scene.
• Aurora gradient shader needs a section-color setter. When you switch slides, call scene.setColor(ACCENT[color]) so the divider's radial gradient matches. Forgetting this is why every section divider looks the same.
• Build-step navigation runs first. Pressing → first runs build steps inside the slide (progressive bullet reveals, code-line highlights), then advances to the next slide once steps are exhausted. Track this in slide controller state, not on the scene.
• Real-data discovery is read-only. Whether your source is kubectl, a SQL query, a REST API, or a file scan — pick read-only operations during deck-building. Never run write / destructive ops while authoring slides.
• Pick a concrete metaphor before coding the scene. "A 3D representation of X" is not a scene — it's a wish. Land on a specific physical metaphor (boxes filling a server, particles flowing through gates, balls bouncing on a curve, a force-directed graph relaxing) before writing Three.js.
• Topic-agnostic does not mean topic-blind. The framework is generic; each deck is specific. Start with the narrative, then the metaphors, then the scenes — not the other way around.

What you DO

1. Treat the framework templates as the source of truth — copy and extend, do not refactor speculatively.
2. Capture real data with the appropriate discovery script before writing the "real numbers" section.
3. Build each scene as a small, focused module (~150-300 lines) illustrating exactly one concept.
4. Use the section accent color in code-block borders, progress bar, kicker text, and aurora gradient — visual continuity matters.
5. Write speaker notes inline in <aside class="notes"> so they show up in both the runtime panel (S key) and the generated SPEAKER-GUIDE.md.
6. Add a ▶ Play full-screen toggle button to every scene slide so the 3D scene can take over without slide content covering it.
7. Pair this skill with complementary skills: a presentations / presentation-design skill for animation principles + slide design, a presentation-content skill for content writing, a threejs-webgl skill for 3D fundamentals, an outline / pitch-deck skill for narrative structure.

What you do NOT do

• Generate Three.js scenes from prose ("a 3D representation of "). Pick a concrete metaphor first, then code.
• Put more than 5 bullets on a slide — split it.
• Use stock images, clip art, or emoji decoration. The deck's visuals are Three.js + accent colors + typography.
• Animate everything — section dividers and big-statement slides should be still.
• Skip the speaker guide. The deck is half the artifact; the talk is the other half.
• Commit credentials, internal hostnames, real customer data, or anything that leaks production / private info beyond what is already public.
• Re-implement Three.js fundamentals in this skill's references — defer to a threejs-webgl skill. This skill teaches the integration pattern (one canvas, scene registry, slide-driven scene switch, accent-color sync) — not how to use materials and lights.
• Make assumptions about the topic. The framework is topic-agnostic; ask the user for their narrative before scaffolding scenes.