PromptBank

Table of contents

• Overview
• Prompt creation workflow
• Feature
• Version diff engine
• Structure
• Technologies used
• Installation
• Contact
• Room for improvement

Overview

PromptBank is a prompt engineering workspace with:

• A Go backend API for auth, prompt management, versioning, and AI-assisted operations
• A Next.js frontend for composing prompts and managing prompt assets
• Gemini-powered helpers for normalization, framework suggestion, scoring, and iterative refinement

The project focuses on making prompt creation structured, repeatable, and versioned.

Prompt creation workflow

PromptBank supports both creating a prompt from scratch and deriving a new prompt from an existing one. Users can enter raw prompt ideas, manually edit structured asset fields, ask the LLM to normalize fuzzy input into reusable assets, select frameworks and techniques, compose the final prompt, test it in the playground, and save the result as a new version.

Feature

• User authentication (register/login) with JWT
• Prompt CRUD and ownership-scoped access control
• Prompt versioning and derive flow
• Prompt composition from:
  • Asset fields (goal, persona, context, tone, constraints, examples)
  • Selected framework
  • Selected techniques
• LLM capabilities:
  • Wizard answer normalization
  • Framework suggestion
  • Prompt quality scoring
  • Iterative refinement with tool-calling
  • Technique-specific AI suggestions (few-shot examples, role persona, constraints)
• Frontend pages for signup/login, prompt list, prompt detail/composition, version browsing

Version diff engine

PromptBank includes a structured diff engine for prompt versions in internal/diff/jsonpatch.go. Instead of comparing composed prompt text, it compares the normalized version document: assets, framework_id, and technique_ids.

The engine emits an RFC 6902-style subset of operations (add, remove, replace) with JSON Pointer paths and summary stats. Ordered arrays such as examples are compared by index, while set-like fields such as technique IDs are compared by membership. The API exposes this through GET /api/v1/prompts/{promptID}/versions/diff?from=X&to=Y, and version creation can cache the parent diff in prompt_versions.diff_from_parent so later lineage and history features do not need to recompute it.

Structure

PromptBank/
├─ cmd/
│  ├─ api/                 # API entrypoint
│  └─ worker/              # Worker scaffold (currently placeholder)
├─ internal/
│  ├─ http/                # Router + handlers
│  ├─ llm/                 # Gemini client + normalize/score/refine/suggestion logic
│  ├─ repository/          # Postgres repositories
│  ├─ security/            # JWT and auth helpers
│  ├─ composition/         # Prompt composing pipeline
│  ├─ framework/           # Framework definitions and slot mapping
│  ├─ technique/           # Technique definitions and apply logic
│  └─ asset/               # Asset models and normalization
├─ db/
│  └─ migrations/          # SQL migrations embedded into the API binary
├─ docs/
│  └─ assets/              # README and documentation assets
├─ frontend/               # Next.js app
├─ docker-compose.yml      # Local infra: Postgres + Redis + API
└─ README.md

Technologies used

• Backend: Go (Chi router, pgx, JWT)
• Frontend: Next.js, React, TypeScript
• Database: PostgreSQL
• Cache/queue-ready infra: Redis
• AI integration: Google Gemini (google/generative-ai-go)
• Dev environment: Docker Compose

Installation

1) Backend + infrastructure

1. Copy environment file:

   cp .env.example .env

2. Start services:

   docker compose up --build

3. Verify API health:

   curl http://localhost:8080/health

2) Frontend

1. Configure frontend env:

   cd frontend
   cp .env.local.example .env.local

2. Install and run:

   npm install
   npm run dev

3. Open:
   • http://localhost:3000

Notes:

• Default frontend API base is /backend (rewrite-based proxy to avoid CORS issues)
• If you change NEXT_PUBLIC_API_BASE_URL or BACKEND_ORIGIN, restart the frontend server

Contact

• Open an issue in this repository for bugs, feature requests, or questions.

Room for improvement

• Async worker pipeline
  cmd/worker is currently a scaffold. Move long-running/expensive AI tasks to background jobs (e.g., Redis-backed queue) and let API endpoints return job IDs + status polling/webhooks.

• Prompt lineage UX and deeper lineage features
  Lineage is stored on derive (prompt_lineage) but can be expanded with:

  • lineage graph view in frontend
  • branch comparison between versions/prompts
  • clearer ancestry metadata in prompt detail and version cards