Skip to content

Semantic Prompt Language (SPL)

An agent-first, strongly-typed functional language with row polymorphism, first-class prompts, compile-time JSON Schemas, and colorless async — compiled cleanly to ES Module JavaScript.
import llm from "std/llm"
import os from "std/os"
/// Developer doc-comment (shown in IDE hover, omitted from LLM schemas).
'' Historical figure profile returned by the model.
type Person = struct {
firstName: Str '' Given name, e.g. Ada
lastName: Str '' Family name, e.g. Lovelace
age: Int '' Age in years
profession: Str '' Primary occupation
}
let api_key = os.env("OPENAI_API_KEY") ?? panic("OPENAI_API_KEY not set")
// Automatically sends `type_schema<Person>()` and validates the JSON response!
let ada = llm.ask<Person>(
provider: "openai_responses",
model: "gpt-6-luna",
api_key: api_key
) @{
Generate a profile for Ada Lovelace.
}?
print("${ada.firstName} ${ada.lastName} (${ada.age}) — ${ada.profession}")

First-Class Prompts & Schemas

Write auto-dedented, brace-balanced @{ … } prompts with ${expr} interpolation, attach ‘’ LLM hints directly to fields, and generate JSON Schemas + decoders at compile time via type_schema<T>() and decode<T>().

Hindley-Milner + Row Polymorphism

Zero null or undefined. Structural records, functional spread ({ …x, k: v }), open records (..r), hybrid dynamic records (..extra: Dict<Str, Json>), and exhaustive match.

Colorless Async & Agents

No async/await clutter. The compiler infers suspension automatically, runs map sequentially to prevent runaway fan-out, and offers opt-in par_map, multi-provider std/llm, and deterministic agent_replay<T>.

Compiles to Clean ESM

Targets Browser and Node 22+ / Bun / Deno with zero root runtime dependencies, self-tail-call optimization, reifiable generics (reified fn), and a built-in LSP server for VS Code.