DEV Community

ShovelMaker
ShovelMaker

Posted on

I got tired of inconsistent AI code, so I built a PRD validator

#ai #api #cloudflare #typescript

I asked Claude to "build a login system" and got different code every time.

Sometimes JWT. Sometimes sessions. Sometimes OAuth appeared out of nowhere.

Got frustrated, so I wrote down exactly what I wanted. Then the output became consistent.

But how do you know if that doc is actually complete?

So I Built RC Gate API

Throw in your PRD markdown, it tells you what's missing.

POST /prd/validate
{"prd_markdown": "## Summary\n..."}
Enter fullscreen mode Exit fullscreen mode

Response:

{
  "score": 65,
  "errors": [
    {"code": "MISSING_SECTION", "message": "Missing: Error Handling"},
    {"code": "TODO_FOUND", "message": "3 TODOs remaining"}
  ]
}
Enter fullscreen mode Exit fullscreen mode

AI Analysis is Paid Only

Running AI on every request gets expensive fast.

So I set it up like this:

Tier Price What you get
Free $0 Rule-based only (section checks, TODO detection)
Pro $9 Adds Claude Haiku analysis
Ultra $29 Claude Sonnet deep analysis

Rules are free. Pay if you want AI.

Tech Stack

  • Cloudflare Workers ($0 hosting)
  • Hono
  • Anthropic API

How is this different from GitHub Copilot?

Copilot reviews your code after you write it.

This validates your spec before you write code.

Different timing.

📄 Docs: https://adspy-api.com/docs

🔧 Swagger: https://adspy-api.com/swagger

Top comments (0)