Hand an AI agent a topic. It researches, designs, writes the copy, generates the art, and ships a self-contained HTML deck — interactive, version-controlled, exportable to PDF/PPT. No PowerPoint required.
Web + your vault; pulls real numbers.
One idea per slide; dark→light→dark.
Generates art + CSS data-viz, no chart lib.
Overflow contract; iterates to fit the frame.
Live HTML, PDF, or PPTX.
Click each panel to reveal the answer and rack up points — exactly the kind of audience game you can drop into a talk. Built in plain JS, runs anywhere.
Static charts freeze an idea. An interactive one lets the audience watch it build. Hit play to animate the model bake-off scores.
Same <deck-stage> template, design system, and image generator — handed to four frontier models with the brief "build a short About Hermes deck." Verdicts are my own visual review; tokens & cost are from the session logs.
The most complete deck — cinematic art, cohesive violet/gold palette, inline mono code, a captioned knowledge-graph render. Felt designed, not generated.
Clever winged-sandal motif, teal-on-violet gradient, tidy 3-card layouts, page footers, a witty "Dramatization" caption. Editorial and disciplined.
Strong winged-helmet title, good split content slide with a real knowledge-graph image. One flaw: pill emoji rendered as tofu boxes.
Pretty title, but the content slide fell apart: tofu icons, cramped labels, misaligned bars. Used every tool yet couldn't land the craft.
Skipped the available tools and looped — bad look-and-feel, off formatting, no deck retained. The ~320K tokens were the loop, not the task.
The split is two skills: tool discipline (did it use the template + image gen?) and design taste. Sonnet & GPT cleared both; GLM got ~90% there; DeepSeek used the tools but cramped the slide; Grok skipped them.
Every one is a live HTML deck — arrow-key nav, animations, print-to-PDF. Click to open.
● Live deckEnd-to-end: the agent pulled live competitor data and built an investor-grade deck — charts, unit economics, a verdict.
● SonnetThe most complete of the bake-off — cinematic art, cohesive palette, a captioned knowledge-graph render.
● GPT 5.5Clever winged-sandal motif, teal-on-violet gradient, tidy layouts, a witty "Dramatization" caption.
● GLM 5.2Strong winged-helmet title and a real split content slide. One trip-up: a couple of emoji rendered as tofu.
● DeepSeekPretty title, but the content slide cramped up — proof that tool-use alone isn't taste.
● Related talkMy most recent talk used this technique — with a hand-coded SVG calibration animation right on the slide.