prompt-optimizer

Abstracts multiple LLM providers (OpenAI, Anthropic, Google Gemini, DeepSeek, SiliconFlow, Zhipu AI) through a unified service layer that handles model configuration, API credential management, and request routing. The system maintains a model registry with provider-specific parameters and implements adapter patterns for each provider's API contract, allowing users to swap models without changing optimization logic. All API calls execute client-side with credentials stored locally in IndexedDB, eliminating intermediate server dependencies.

Implements a template system that defines optimization workflows as reusable patterns with placeholder variables. The system automatically extracts variables from user input using regex and semantic analysis, then applies templates through a substitution engine that generates optimized prompts by filling placeholders with extracted values. Templates are stored as configuration objects with metadata (name, description, category) and can be customized per-user or shared across workspaces. Variable extraction uses both pattern matching and LLM-assisted detection to identify dynamic content.

Implements comprehensive internationalization (i18n) across all platforms with support for English, Chinese (Simplified and Traditional), and other languages. The system uses Vue.js i18n plugin with locale-specific message files, supports dynamic language switching without page reload, and maintains language preference in local storage. UI components are designed to handle variable-length text across languages, and all user-facing strings are externalized from code.

Implements a VCR (Video Cassette Recorder) testing system that records and replays HTTP interactions with LLM provider APIs, enabling deterministic testing without live API calls. The system captures request/response pairs during test execution, stores them as YAML cassettes, and replays them in subsequent test runs. This approach eliminates API rate limiting issues, reduces test latency from seconds to milliseconds, and enables testing without valid API credentials. Cassettes are version-controlled alongside test code for reproducibility.

Provides containerized deployment through Docker with environment variable configuration for API credentials, model settings, and feature flags. The system includes Docker Compose configuration for local development and production-ready Dockerfile for container registry deployment. Vercel deployment is configured through vercel.json with automatic builds and deployments on git push. Environment variables are externalized from code, enabling secure credential management across deployment environments without code changes.

Implements application state management using Pinia (Vue.js state management library) with reactive stores for prompts, models, templates, and user preferences. The system persists state to IndexedDB on every change, enabling automatic recovery on page reload or application restart. Pinia stores provide centralized state access across all components, with computed properties for derived state and actions for state mutations. Session state includes active workspace, selected models, and UI preferences.

Enables users to export prompts, templates, and workspace configurations in JSON format and import from external sources with format validation. The system implements schema validation to ensure imported data matches expected structure, performs data migration for version compatibility, and provides detailed error reporting for invalid imports. Export includes full metadata (timestamps, optimization history, evaluation results), and import can merge with existing data or replace it entirely. Supports batch import/export for multiple workspaces.

Enables users to conduct multi-turn conversations with multiple LLM models simultaneously, displaying responses in a multi-column layout for direct comparison. The system maintains conversation history per model, tracks token usage and latency metrics, and allows users to branch conversations at any turn. Each model maintains independent state and context windows, with the UI rendering responses in synchronized columns to highlight differences in reasoning, tone, and accuracy. History is persisted locally in IndexedDB with full conversation replay capability.

Provides a pluggable evaluation system that allows users to define custom metrics (accuracy, relevance, tone, length, etc.) and apply them to prompt optimization results. The evaluation engine executes metrics as LLM-based judges or rule-based scorers, aggregates results into a scoring framework, and ranks optimized prompts by composite scores. Metrics are configurable per evaluation run, with support for weighted scoring and threshold-based filtering. Results are persisted with full audit trails for reproducibility.

Extends prompt optimization to include image inputs, allowing users to upload images and generate or optimize prompts that reference visual content. The system integrates vision-capable LLM models (OpenAI GPT-4V, Google Gemini Vision, Claude Vision) to analyze images and suggest prompt improvements that account for visual context. Images are stored locally in IndexedDB with base64 encoding, and the optimization engine maintains image-prompt associations throughout the optimization workflow. Supports batch image processing and multi-image prompt generation.

Organizes prompts into workspaces with mode-specific optimization strategies: Basic Mode (simple prompt refinement), Pro Mode (advanced optimization with context management), and Context Mode (full conversation context integration). Each workspace maintains isolated state, configuration, and history. The system implements mode-specific UI components and optimization logic, allowing users to switch between strategies based on task complexity. Workspaces are persisted in IndexedDB with full import/export capability for sharing and backup.

Maintains a complete version history of all prompt modifications with branching capability, allowing users to explore alternative optimization paths and revert to previous versions. The system tracks changes as immutable history entries with metadata (timestamp, model used, optimization type), implements a branching model where users can create alternative versions at any point, and provides replay functionality to re-execute optimization steps. History is persisted in IndexedDB with full diff visualization between versions.

Allows users to mark prompts, templates, and optimization results as favorites with custom tagging and categorization. The system implements a favorites store in IndexedDB with metadata (creation date, usage count, rating), supports full-text search across favorites, and provides smart organization through auto-generated categories based on content analysis. Favorites can be exported as shareable collections and imported from external sources, enabling prompt sharing across teams.

Distributes the same core optimization engine across three deployment platforms: web application (Vercel), browser extension (Chrome/Firefox), and desktop application (Electron). The system uses a monorepo structure with shared UI components (Vue.js), core services, and platform-specific adapters. Each platform maintains local IndexedDB storage for prompts and configuration, with optional cloud sync capability. The architecture enables users to access their prompts across devices while maintaining client-side processing and local data storage.

Integrates with the Model Context Protocol (MCP) to expose prompt optimization capabilities as a server that external tools and agents can invoke. The system implements MCP resource and tool definitions that allow other applications to call optimization functions, retrieve prompt history, and manage workspaces through a standardized protocol. MCP server runs locally alongside the main application, enabling seamless integration with AI agents and automation workflows without API gateway complexity.