Models
Back to ideas
Agent · explainer

On-demand visual explainer

Pick any frontier-AI concept. A model writes a 5-6 slide swipe-through with structured visuals on the fly — same Story format we use for the hand-built explainers, generated end-to-end by an agent. The visuals come from a fixed primitive library (callouts, stat pairs, sequences, comparisons, bars), not arbitrary code, so the output stays safe and consistent.

How it works
  • Schema-constrained generation. The model emits structured JSON, not free-form HTML or React code. Each slide picks a visual type from a fixed list (callout, stat-pair, stat-stack, sequence, before-after, bar-compare).
  • Deterministic rendering. The visual primitives are normal React components. There is no eval, no sandbox, no arbitrary code execution.
  • Premium voice. The system prompt enforces Pere's editorial stance — specific numbers, no hedge endings, no marketing copy, no “dive in” openers.
  • Honest failure. If the model returns malformed JSON or fewer than 3 valid slides, the request fails loud rather than rendering a half-broken story.