deck-visuals

deck-visuals

Your deck already reads. Make it land.

deck-visuals is a Claude Code plugin that decides where a visual belongs in an HTML deck and which kind, then hands back a ready-to-inject fragment. Inline-SVG concept diagrams, charts with a validated palette, icon anchors, and GIFs it checks before they ship. It does not render decks.

Four bullets in, one chart out. The numbers are the same; the argument lands.

See it live: gallery · showcase · how it decides · site

Install

/plugin marketplace add gagoar/gago-plugins
/plugin install deck-visuals@gago-plugins
/reload-plugins

The three pillars

1. Impactful visuals. Inline-SVG concept diagrams for relationships, flows, and architecture. Icon anchors for single ideas. AI illustrations for hero moments are deferred — this build has no image-generation dependency to satisfy that case yet.
2. Data-viz principles. Charts lean on the built-in dataviz skill for form, marks, and interaction, paired with a validated brand palette instead of dataviz's brand-neutral default. Every chart ships with a reachable table view.
3. Levity. Real GIFs by direct media URL, at a deliberate cadence — openers, section breaks, the payoff — with alt text and a prefers-reduced-motion fallback on every one. A swipe onboarding flow rebuilds the presenter's pop-culture profile so the matcher always has options, backed by a meme library for when nothing else fits.

Guidance and injection, not rendering

frontend-slides renders decks: the HTML shell, the fixed stage, the theme, the animations. deck-visuals never touches that layer. It reads a deck's content, decides where a visual belongs, builds that one visual as a self-contained fragment with its own scoped styles, and hands it back for injection into the slide that's already there. iceberg still owns the copy pass, run on prose before it enters the deck — never on the finished .html file.

See skills/deck-visuals/SKILL.md for the full five-step pipeline and skills/deck-visuals/references/ for the per-flavor rules.

Usage

Ask in natural language once a deck exists:

• "Add visuals to my deck."
• "Make this slide impactful."
• "Chart this number."
• "Add a GIF/levity slide here."
• "Inject a diagram for this architecture slide."

The skill reads the deck, applies the decision rule in skills/deck-visuals/references/visual-concepts.md, and returns the fragment to inject — it does not rewrite the deck's shell or theme.

What's inside

skills/deck-visuals/
  SKILL.md                    orchestrator: pipeline + trigger phrases
  references/                 decision rules, per-flavor guidance
                              (incl. onboarding.md)
  assets/
    brand-palette.md          validated 8-slot categorical palette + ramps
    show-catalog.md           the ~50-title swipe catalog (onboarding)
    meme-library.md           evergreen meme fallback for levity
    familiar-sources.md       the presenter's curated analogy pool
    VERSION                   release tag the wrappers download from
    run-binary.sh             shared fetch + verify + cache + build-fallback
    tools/                    dv-tools: Go source for the palette validator,
                              media-URL check, and coverage report
      dv-tools                POSIX wrapper: fetches the pinned binary for this
                              machine on first use, then runs it
      SHA256SUMS              pinned hashes of the released dv-tools binaries
      build.sh                reproducible cross-compile; rewrites SHA256SUMS
    icons/                    curated Lucide-style icon set
    fragments/                injectable HTML fragments
    onboarding/
      picker.html             onboarding swipe page (local tool)
      levity-picker.html      deck-time levity GIF chooser (local tool)
      palette-picker.html     accent-theme chooser (local tool)
      picker.css              shared styling for all three: the site's cream and orange look
      fonts/                  Fraunces + IBM Plex, compiled into the binary (OFL licenses inside)
      dv-onboard              POSIX wrapper for the picker binary (same scheme)
      SHA256SUMS              pinned hashes of the released dv-onboard binaries
      server/                 Go source for the picker binary

The same picker server also powers a deck-time levity picker: once Claude proposes 2-3 candidate GIFs per slide, run it with --form levity to choose each one visually instead of in chat.

Helper binaries

dv-tools (palette validator, media-URL check, coverage) and dv-onboard (the picker server) are Go programs. Their binaries are not committed. The wrappers (assets/tools/dv-tools, assets/onboarding/dv-onboard) download the binary for the installed release on first use, run it only if it matches the pinned SHA256SUMS, and cache it. They need network access to GitHub, or Go installed the first time; after that they work offline. Maintainers: bump assets/VERSION, .claude-plugin/plugin.json and the SHA256SUMS files (run both build.sh scripts) together, then push a v<version> tag so release-binaries.yml publishes the assets.

Onboard your references

"Onboard me" / "update my familiar sources"

Swipe through ~50 shows in a local page — each a still plus a signature GIF — marking what you recognize, and give your talk length and audience. The picks feed back through a small local-only binary (nothing leaves the machine) and rebuild familiar-sources.md, with a coverage check that proves the matcher has ~3 options per concept shape.

License

MIT. See LICENSE. The icon set carries its own provenance note and license text in skills/deck-visuals/assets/icons/LICENSE.

Automated gate

CI runs scripts/check-plugin-safety.ts and claude plugin validate --strict on every PR. Run both before you push:

node scripts/check-plugin-safety.ts .
claude plugin validate . --strict

The gate fails on three things the plugin-directory upload scanner holds for review:

• A skill, agent, or command that pre-approves Write, Edit, or Bash with no path scope. Write Write(./docs/**), not Write.
• Shipped code that copies the whole environment (...process.env, Object.keys(process.env)). Read the named variables you need.
• A committed native executable. The scanner cannot read it.