I posted an early version more than 20 days ago, but the system has changed enough that it feels like a different product.

Originally, users had to type detailed text instructions (“text to form”) to generate a form, document template, and API. It worked, but people weren’t enthusiastic about writing formatted text.

The new version removes that step entirely. The user now provides only a name or intent (“medical intake form”, “sales contract without required fields”, etc.). Genie Snap constructs a strict, rule based prompt, and the LLM’s only job is to output a block of formatted text that follows those rules (field syntax, list rules, naming conventions, etc.).

Genie Snap then deterministically compiles that formatted text into: • a form • a document template • a REST API

The LLM never generates workflow logic — it only fills a structured output contract. The compiler enforces all structure and behavior, which makes the process repeatable and avoids hallucinations.

A human review step is built in before publishing, since the LLM’s wording can vary between runs even though the structure is always valid. Once a workflow is approved, it becomes a deterministic, repeatable self service process.

Genie Snap also hosts the workflow and manages submissions, so the forms and APIs run immediately without additional infrastructure.

Everything works fully offline (Ollama optional) on standard CPU only hardware.

So far it has handled over 1,000 generations without structural failures. I’d love feedback from anyone interested in deterministic systems, compilers, or LLM assisted tooling.

Link: https://www.geniesnap.com

How it works (technical overview) Genie Snap doesn’t use the LLM to generate workflow logic. Instead, it gives the LLM a strict formatting contract: field rules, list rules, naming conventions, ASCII only, no commentary, etc. The LLM’s only job is to output a block of formatted text that follows those rules.

Genie Snap’s compiler then deterministically generates the form UI, document template, and REST API from that formatted text. The compiler enforces all structure and behavior, so the workflow generation is repeatable and doesn’t depend on LLM creativity.

The LLM’s wording can vary between runs, but the structure is always valid. A human review step exists before publishing. After approval, the workflow becomes a deterministic, repeatable self service process.