PROMPT SYSTEM

# Prompt system architecture

> The model does not execute templates. It only responds to fully resolved input.

This document captures the mental model, architecture, and build principles for how PromptLab treats prompts as compiled artifacts rather than instructions sent raw to an LLM.

---

## Core principle

The LLM is a pure function at the end of your pipeline — not the pipeline itself.

Everything that feels "smart" — variable handling, data binding, workflow logic, validation, bullet enforcement — is your responsibility upstream. The model's only job is formatting and tone enforcement on clean, validated, fully resolved input.

If the model is doing gap-filling, inferring missing fields, or producing inconsistent output, the problem is almost always in layers 1–4, not in the prompt itself.

---

## The 5-layer pipeline