Web Dev Mehedi：因厭倦含糊 PRD，開發可同步產出架構的生成器

Every project I start the same way: a PRD from a chatbot, good intentions, and then — three days into building with an AI coding agent — a database that doesn't match the document, an API payload that changed shape twice, and no security rules anywhere.

The problem wasn't the agent. It was my spec. Prose is a suggestion; a schema is a contract.

So I built the tool I needed. Web Dev Mehedi takes one product idea and generates four files that agree with each other:

PRODUCT_REQUIREMENTS.md — scope, user journeys, edge cases, and GIVEN/WHEN/THEN acceptance criteria.
UI_UX_SPECIFICATION.md — design tokens, typography, view states, accessibility rules.
TECHNICAL_ARCHITECTURE.md — PostgreSQL DDL schemas with Supabase Row Level Security, plus typed REST & Server Actions API contracts with Zod validation.
AI_AGENT_IMPLEMENTATION.md — phased build instructions and prompt rules that pin Cursor and Claude Code to files 1–3: use these exact table names, these exact payloads, never bypass RLS, and flag conflicts instead of choosing silently.

Three lessons from building it:

Multi-tenancy has to be in the schema on day one, not patched later. RLS policies are written in the same step as the tables themselves.
An agent follows rigid rules surprisingly well. Most "hallucinations" I saw were vagueness failures, not capability failures: the spec never actually decided the thing the agent got wrong.
Pricing in BDT with bKash opened a market that $20/month card subscriptions simply don't reach. Web Dev Mehedi starts at ৳99/month, with 50 free credits on signup and no card required.

It's live here if you want to break it: https://webdevmehedi.com/ai-prd-generator — and the free template behind the format is on GitHub: https://github.com/web-dev-mehedi/ai-prd-template

What do you hand your coding agent before a build? Genuinely curious what other people's specs look like.