PDD (Prompt-Driven Development)

PDD is an Alpha-stage (v0.0.216) AI development tool maintained by Prompt Driven, Inc. Its core philosophy comes from the PDD Doctrine—treating Prompts and tests as "molds" and code as "produced goods," emphasizing the compounding effects of Test Capital, Prompt Capital, and Grounding Capital.

Driving Modes#

Issue-Driven Mode: Multi-step Agentic workflows automatically produce code, tests, or architecture files directly from GitHub Issues:

• pdd change <issue-url> — 12-step workflow to implement feature requests
• pdd bug <issue-url> — Create failing tests for bugs
• pdd fix <issue-url> — Fix failing tests
• pdd split <target-file> — 14-step workflow to diagnose and split large development units
• pdd generate <issue-url> — Generate architecture.json from PRD Issues (11 steps)
• pdd test <issue-url> — Generate UI tests with exploratory testing and accessibility audits (18 steps)

Prompt-File-Driven Mode: Using .prompt files as the authoritative source:

• pdd sync <module> — Fully automated development cycle (intelligent decisions + real-time visual feedback)
• pdd generate <prompt-file> — Generate code from prompt files

Runtime Capabilities#

• Web UI: pdd connect launches localhost:9876, supporting command execution, file browsing, and PDD Cloud remote access
• Multi-LLM Provider: Unified interface via LiteLLM supporting OpenAI / Anthropic / Google / Vertex AI, etc.
• Agentic CLI Integration: Delegates to Claude Code, Gemini CLI, or Codex CLI as execution engines
• Auto Model Selection: Built-in CSV model catalog (ELO ≥ 1400), auto-optimized by coding ability, with user-level and project-level CSV override support
• Cloud/Local Dual Mode: Default cloud mode (GitHub SSO, no local API key needed), --local for pure local mode
• Auto-Update: Built-in auto-update mechanism

Architecture Highlights#

Built on Python (≥3.12) with Click CLI framework, embedded FastAPI + Uvicorn + WebSockets web server, and frontend assets bundled in pdd/frontend/dist. Core logic resides in pdd/core/, command implementations in pdd/commands/. Each workflow (change/bug/fix/split/test/sync/architecture, etc.) has its own agentic_*.py agent and agentic_*_orchestrator.py orchestrator, forming a multi-step serial orchestration pattern. Prompt templates are stored as *_LLM.prompt files in pdd/prompts/.

Key dependencies include: litellm[caching] (multi-provider routing and caching), z3-solver (constraint solving and structured output validation), GitPython (repository operations), firebase_admin + keyring (cloud authentication), firecrawl-py (web content extraction), watchdog (file change monitoring), etc.

Installation & Usage#

curl -LsSf https://astral.sh/uv/install.sh | sh
uv tool install pdd-cli
pdd setup

Issue-driven mode requires GitHub CLI (gh) and at least one Agentic CLI (Claude Code / Gemini CLI / Codex CLI) pre-installed.

Unconfirmed Information#

• Independent website URL and PDD Cloud service details (hosting, pricing, SLA) are not clearly specified
• Whitepaper link mentioned in README (with benchmarks and case studies) was not fully captured
• Complete schema/DSL specification for .prompt files remains to be confirmed
• Windows compatibility and minimum Agentic CLI version requirements are not explicitly documented