Prompting today is like building with sand.

2023 · at VoiceflowIn 2023, at Voiceflow, we wrote our first prompt. You typed a sentence in plain English, no syntax, no ceremony, and working software answered back.

It was magic. Everything we knew about programming said it shouldn’t work. It worked anyway.

So we built. Chatbots, then agents, then whole products with a paragraph of English where the code used to be. The prompts grew: a sentence, then a page, then a folder of pages.

Somewhere in there, the thing under our hands changed. The model stopped being a chat window. It had become a computer. A strange new kind of machine, one that runs English.

Which means the prompts inside our products were never really “prompts.” A user prompt is a question: you ask it once, it answers, and it’s gone. An application prompt ships. It runs ten thousand times a day, unattended, in front of your customers. An application prompt is a program.

And at Voiceflow, we had shipped those programs everywhere - thousands of agents on support desks, sales floors, systems people counted on. Every one of them balanced on prose.

Then, a few years in, it started to feel like building with sand.

We’d shore up one behavior here, and something would collapse over there.

Change one line? Broken.Swap the model? Broken.Trim it for tokens? Broken.Reword a sentence, move a paragraph, add a comma…Broken. Broken. Broken.

And nothing told us why. No error, no line number, no red squiggle. Just a customer, three days later, finding the one answer the demo never showed.

Because look at what we were actually doing: writing programs with no compiler, no types, no syntax to hold a shape. We were writing code by hand, in Notepad, and shipping it to production on hope. The most important programs in the company were the only ones nothing checked.

We’d seen this movie before. JavaScript felt like this once: fast, loose, terrifying at scale. Then TypeScript gave it a spine. The same freedom, with a net underneath.

English deserved the same.

A language where personas are typed and “warm” is a contract, not a vibe. Where MUST means must, tests live beside the words they verify, and a checker proves the whole thing can’t contradict itself, before a single user ever sees it.

Sky MUST offer @{upgrade}
Sky NEVER upsells
// E201: these cannot both hold

$DICTIONARY "./brand.dict"
Lexicon MUST use "sign in", NEVER "log in"
Style NEVER uses "-" // yes, even that

Persona MUST cite {policy}
  · guards 3 tests · fired 214×/day
Persona SHOULD stay brief
  · overlaps line 12 · safe to delete

$STATE checkout
  SHOW rules FOR payments
  Tools ARE refund, invoice // 2 of 41

$CONFIG model claude-opus-4-8
$CONFIG model gpt-5.2 // same file, recompiled

$RULES "./team.rules.tg"
agent: draft line 14 rejected (R7)
agent: patched · recompiled ✓

The sand is still sand. That’s the point. It’s still English, still fast, still magic. TypeGlish is the form that makes it hold.