explorable library · recipe

Build an essay

An essay is long-form writing where the prose itself is interactive: change a number, or your whole point of view, and the argument re-reasons around you. Here is the recipe, and a prompt you can hand straight to an AI.

Anatomy

  1. Put one coupled model under the whole piece — a compute() the entire essay reads from, fed by every scrubbable number and every chart handle, so changing one re-reasons the prose, the charts and the conclusions at once, possibly reframing the entire essay — which is the point.
  2. Make the prose reactive: sentences rewrite at thresholds, reasoning backward to the cause and forward to the consequence, with the words changing as the reader drags.
  3. Keep callouts permanently live: recompute their numbers on every change and never freeze them at the values they had when first triggered.
  4. Build a learning layer (especially for short or guided pieces): open with a question, nudge the first move, give a payoff that finds the reader (a callout that appears and names the insight), and make the insight a place the reader reaches for and discovers they cannot have.
  5. Never assert a floor or threshold — decompose it so the reader sees exactly why it exists.
  6. Optionally add a perspective lens (worker, employer, treasury, society) that reframes the whole essay and reorders what matters.
  7. Stay honest: label model versus fact clearly, mark illustrative forecast tails, never fabricate, and keep motion steady with no flashing and reduced-motion gated.

The build prompt

Copy this into your agent:

You are building a reactive explorable ESSAY about {TOPIC} — long-form writing where the prose itself is interactive.

1. One coupled model under the whole piece: a compute() the entire essay reads from. Every scrubbable number in the sentences AND every chart handle feeds it; change one and the prose, the charts and the conclusions all re-reason. Changing things may reframe the entire essay — that is the point.
2. Reactive prose: sentences rewrite at thresholds (reason backward to the cause, forward to the consequence). The words change as the reader drags. Keep callouts PERMANENTLY LIVE — recompute their numbers on every change; never freeze them at trigger-time values.
3. The learning layer (especially for short/guided pieces): open with a QUESTION; nudge the first move (try this ->); a payoff that FINDS the reader (at the meaningful state a callout appears and names the insight); and make the insight a PLACE the reader reaches for and discovers they cannot have. Do not assert a floor/threshold — DECOMPOSE it so the reader sees why.
4. Optional but powerful: a PERSPECTIVE lens (e.g. worker / employer / treasury / society) that reframes the whole essay and reorders what matters.
5. Honest: model vs fact clearly labelled; illustrative forecast tails marked; never fabricate.
6. No flashing; steady motion; reduced-motion gated.

Kill it if: you can twiddle but it is not obvious what you would learn; a load-bearing number is asserted instead of explained; a callout shows stale numbers; or any value is fabricated.

Kill-criteria

  • You can twiddle but it is not obvious what you would learn.
  • A load-bearing number is asserted instead of explained (decompose it).
  • A callout shows stale numbers instead of recomputing live.
  • Any value is fabricated.

See it in the wild