Prompt Studio

01 Problem

Working with LLM prompts as ad-hoc text makes them hard to reuse, version, and run consistently. Comparing the same prompt across models and providers, keeping reusable prompt components organized, and exporting results all end up scattered across chat windows and scratch files — with no single place to compose, run, and evaluate.

02 Solution

A browser-based prompt studio: a SvelteKit "studio" frontend backed by a FastAPI "builder" that composes prompts from reusable components, runs them across providers through an LLM gateway (OpenRouter), and compares/evaluates the outputs. Layered architecture (settings · llm-gateway · prompt · store · render) composed into modules (prompt-builder · evaluator · tools) with guest and admin storage, auth, and a public-model gate. Deployed on the SBX Contabo pipeline; built with Claude Code CLI under SBX framework governance.

03 Outcomes

04 Constraints

• Prompt components are reusable + composable — not throwaway chat text
• Provider-agnostic model access through a single LLM gateway (OpenRouter)
• Guest storage + admin auth with a public-model gate for anonymous use
• Studio is the only public-facing service; the builder backend is internal-only

05 Innovation

Layered prompt-composition system — settings, LLM gateway, prompt, store, and render layers composed into higher-level modules (prompt-builder, evaluator, tools) so the same primitives back both the product and reusable framework packages. A single LLM gateway abstracts providers so one composed prompt runs and compares across models without rewrites.