Overview

Prompt Development Kit (PDK) is a cross-platform desktop application for building, organizing, and managing complex AI prompt workflows. It provides a visual node-based editor where users can create prompt sequences, manage variables across scopes, and integrate code context from their projects directly into prompts.

The application bridges the gap between simple prompt copy-paste workflows and programmatic prompt engineering, offering a visual interface for prompt development without requiring coding knowledge.

Screenshots

Problem

Prompt engineering for LLMs often involves:

• Managing multiple prompt variations and versions
• Reusing common patterns across different prompts
• Incorporating code snippets and documentation into prompts
• Tracking which combinations work best

Existing solutions force users into either plain text files (no structure) or full programming (high barrier). PDK provides a middle ground: visual prompt composition with variables, versioning, and code integration.

Approach

Stack

• React + TypeScript - Type-safe component architecture for the complex UI
• ReactFlow - Node-based visual canvas for connecting prompt sequences
• Tauri + Rust - Native performance and small binary size for desktop distribution
• SQLite - Embedded database for persistent storage without external dependencies
• Zustand - Lightweight state management for multi-pane layout persistence

Challenges

• Variable resolution order - Implemented recursive variable expansion with cycle detection, distinguishing between content variables (expand recursively) and response variables (literal substitution)
• Multi-pane state persistence - Built a recursive tree structure for pane splits with localStorage persistence, allowing users to view multiple flows simultaneously
• Code context synchronization - Used the notify crate for file watching to auto-update code context variables when source files change

Outcomes

The application successfully handles complex prompt workflows:

• Visual node connections map to prompt sequences
• Three-tier variable system (global, flow, code context) covers common reuse patterns
• Version history enables experimentation without losing previous work
• Token counting provides cost awareness before execution

Key technical wins:

• Cold start under 500ms due to Tauri’s lightweight runtime
• SQLite handles thousands of nodes without performance degradation
• Parallel file scanning via Rayon processes large codebases efficiently

Implementation Notes

The variable system uses a `` syntax with three scopes:

// Variable resolution priority
1. Code context variables (auto-generated from file scans)
2. Flow-specific variables (scoped to current flow)
3. Global variables (available everywhere)

The database schema includes 16 tables tracking flows, nodes, variables, versions, and soft-deleted items:

-- Core tables for flow/node versioning
CREATE TABLE flow_versions (
  id INTEGER PRIMARY KEY,
  flow_id INTEGER,
  version TEXT,
  snapshot TEXT,  -- JSON snapshot of flow state
  created_at TEXT
);

CREATE TABLE node_versions (
  id INTEGER PRIMARY KEY,
  node_id INTEGER,
  version TEXT,
  snapshot TEXT,
  created_at TEXT
);

The code context builder scans project directories and generates variables:

// Parallel file scanning with Rayon
files.par_iter()
    .filter(|f| matches_patterns(f, &include_patterns, &exclude_patterns))
    .map(|f| read_and_tokenize(f))
    .collect()

Related Posts

No posts yet.