arpingblue/vibepm

---

A Staff-level Architect AI Skill designed specifically for Vibecoding (AI-assisted programming). It transforms your vague app ideas into rigid, context-safe technical specifications and step-by-step implementation plans that coding agents (like Cursor or Windsurf) can execute blindly without hallucinating.

🚀 What is this?

The biggest disaster in AI coding is Context Collapse & Stack Hallucination. When you describe a massive app to Cursor all at once, it outputs spaghetti code, mixes up App and Pages router, fails database connections, and gets confused about component states.

VibePM (Vibecoding Architect) solves this by acting as the ultimate planning bridge for execution agents like get-shit-done (GSD) or Cursor. It does not write the app code, nor does it write business plans. It interrogates you before any code is written, locking down 9 core technical elements, aggressively slicing off feature creep, and finally generating native GSD-compatible execution documents (PROJECT.md, ROADMAP.md, and an XML-based PLAN.md).

✨ Core Features

• 🛡️ Extreme Hallucination Prevention: Actively fights scope creep. Whenever a request exceeds the context window of a single AI coding session, it forcibly enforces an MVP cut.
• 🧠 Forced Stack Lock: Forces you to establish the exact Database, UI Library, and State Management strategy before generation. No ambiguous requirements allowed.
• 📄 4 Native AI Artifacts: Generates highly structured PROJECT.md, ROADMAP.md, a strictly XML-formatted PLAN.md with execution waves, and raw Database Schemas.
• 🧩 100% Native: Built purely on Markdown. Loads seamlessly into Claude Code, OpenClaw, Antigravity, or directly into a Cursor Composer context list.

📂 Directory Structure

vibepm/
├── SKILL.md                          # ⚡ Core Architect Engine (<10KB)
├── references/                       # 📚 Anti-Hallucination Knowledge Base
│   ├── discovery-framework.md        # The 9 Vibecoding Technical Elements
│   ├── questioning-rules.md          # Interrogation rules & scope-cutting scripts
│   └── output-templates.md           # 4 templates (.cursorrules, plans, schemas)
├── examples/                         # 💡 Live Examples
│   ├── example-vague-idea.md         # Example of aggressive scope cutting
│   └── example-output-project-brief.md # Example generated plan for an AI
├── .agents/workflows/vibepm.md     # 🔌 Antigravity / OpenClaw adapter
├── .claude/commands/vibepm.md      # 🔌 Claude Code instruction adapter
└── .codex/skills-index.json          # 🔌 Codex / VSCode plugin index

🛠️ Usage / Installation

Load this directory into your preferred AI Agent environment:

For Claude Code

Type the following command:

/vibepm

Or simply say: “I’m going to build a full-stack project with Cursor. Help me architect it using VibePM.”

For Antigravity / OpenClaw

Trigger the workflow by saying:

Run the vibepm workflow. I want to build an AI YouTube summarizer app.

🧐 How it Works

In every interaction, the Skill strictly evaluates the gaps in the 9 Technical Elements:

1. Core Concept
2. Tech Stack Lock
3. Data Schema
4. App Flow
5. State Management
6. External Integrations & Auth
7. Edge Cases
8. Deployment
9. MVP Context-Cut (The most critical element)

Only when these 9 points are absolutely clear and within context limits will it say “Go” and output the hand-off documents for your Coding AI.

🤝 Contributing

If you have highly optimized .cursorrules templates or better prompt architectures for LLM programmers, please open an issue or submit a PR. Our goal is to create the ultimate safety rails for AI developers. For details, see Contributing Guidelines.

---

🌟 Star History

---

📜 License

This project is licensed under the MIT License.