Try It: earl-mcgowen.com/prompt-search

GET /api/prompt-search?q=...&k=5

Overview

This project implements a local-first prompt intelligence system that transforms a GitHub repository of system prompts into a searchable, reusable, and agentic AI asset.
This started as a simple experiment:

“What if prompts weren’t just text… but a dataset you could query, rank, and learn from?”

What came out of it is something much more powerful:

  • A semantic search engine for prompts
  • A Flask API integrated into my AI gateway
  • A SvelteKit frontend for exploration
  • Fully powered by my local GPU using Ollama

The Dataset: Prompt Engineering in the Wild

The foundation of this system is an open GitHub repository:

system-prompts-and-models-of-ai-tools (GitHub)

This repository has become one of the most widely circulated collections of real-world AI system prompts:

  • 120k+ stars and 30k+ forks, making it one of the largest prompt collections on GitHub
  • Frequently referenced as a “goldmine” for understanding how LLMs are guided in production
  • Used to study real system prompts behind modern AI tools

Across the AI community, repositories like this are often treated as:

A “Rosetta Stone” for understanding how AI systems are designed to behave

What makes this particularly valuable is that these prompts reflect:

  • Real production constraints
  • Safety rules and guardrails
  • Tool usage patterns
  • Company-specific prompt design philosophies

The repository includes:

  • System prompts from tools like Cursor, Claude, and VSCode agents
  • Instruction templates and reusable patterns
  • Real-world implementations of AI behavior design

From Repository → Dataset

What stands out is not just the content, but the structure.

Across hundreds of files, clear patterns emerge in how systems:

  • Define roles and constraints
  • Structure reasoning steps
  • Guide tool usage
  • Enforce consistency and safety

These are not isolated prompts. They are designed systems of control and interaction.

Treating this repository as a dataset enables:

  • Semantic search (by intent, not keywords)
  • Cross-tool comparison of prompt strategies
  • Pattern extraction across multiple implementations
  • Generation of new prompts based on proven structures

This insight led to the development of a semantic search engine and prompt generation system built on top of this dataset.

What Was Built

This system functions as a Prompt Intelligence Platform.

Semantic Search Engine

  • Queries prompts by meaning rather than keywords
  • Uses cosine similarity over embeddings
  • Returns ranked, relevant prompt snippets

This is a form of RAG (Retrieval-Augmented Generation) applied to prompts instead of documents.

Prompt Generation

The system extends beyond retrieval.

It enables prompt synthesis.

Prompt Synthesis Flow

User Query: "build a coding agent prompt"
  ↓ retrieve top 5 similar prompts
  ↓ extract patterns
  ↓ synthesize
Output: A structured, production-ready system prompt

This transforms the system into a:

Prompt Generator powered by real-world prompt data

Use Cases

This system enables higher-quality prompt development by shifting from intuition-based design to data-driven prompt engineering.

Instead of isolated experimentation, prompts can be analyzed across contexts, compared structurally, and refined into reusable patterns.

It supports:

  • Standardization of prompt design across teams
  • Creation of reusable prompt templates
  • Faster iteration through semantic comparison
  • Continuous learning from accumulated prompt data

Over time, this evolves into a learning system, where each new prompt builds on prior knowledge rather than starting from scratch.

Industry Applications

Healthcare

  • Clinical decision support prompts
  • Medical summarization prompts
  • Patient interaction prompts

Legal / Attorneys

  • Case summarization prompts
  • Contract analysis prompts
  • Legal research prompts

What This Project Represents

This project reframes prompt engineering as a data system problem.

Instead of treating prompts as isolated artifacts, they are modeled as structured data with:

  • Lineage
  • Embeddings
  • Queryability
  • Measurable relevance

This enables a pipeline similar to modern data systems:

  • Raw prompts → embedded representations
  • Queries → retrieval + ranking
  • Outputs → generated prompts

It also introduces a feedback loop:

Running Everything Locally

The entire system runs on my home setup:

  • Ollama (embeddings + LLMs)
  • SQLite vector database
  • Flask API gateway
  • SvelteKit frontend
  • Exposed via ngrok / (soon Cloudflare Tunnel)

This gives me:

  • Full control
  • No API costs
  • Private data processing
  • Production-like architecture

Here’s your final blog post in clean copy-paste markdown, with your generated hero image included and everything aligned to your site structure.

  • New prompts are added
  • Performance can be evaluated
  • The system continuously improves

What begins as a static repository becomes a living prompt system—capable of evolving over time.

Final Thoughts

Prompt engineering is often treated as a short-lived, trial-and-error process.

At scale, it becomes something else entirely.

When prompts are structured, stored, and analyzed collectively, they become:

A dataset. A system component. A layer of intelligence.

Patterns emerge. Design principles become visible. Effective strategies can be reused and refined.

The result is a new class of systems:


They **learn from them.**