Letting AI code while I do QA?
I am testing a structured workflow for AI-assisted dev.
I’ve been in webdev for years, wearing many hats—webmaster, UI/UX designer, full stack developer, QA, DevOps, tech SEO, and project manager.
That also means I see myself as a crappy coder—but one who understands the full architecture of a developed app. That’s why I’ve been exploring coding assistants to help me create apps from scratch. My assumption? Let the AI handle the coding while I focus on QA.
Directory Structure
To keep things structured, I’ve set up a dedicated .ai/ directory—a context hub for AI-assisted development. It holds key docs like:
Dev Principles & Constraints. It could hold your overarching guidelines, brand voice, code style preferences, etc.
- Philosophical Tenets
- “Always prioritize user privacy.”
- “Respect the user’s time with fast load times.”
- Tech Tenets
- “Keep dependencies minimum to reduce bloat.”
- “Follow TDD at all times”
- “Follow this coding style: _code_style_ “
- Design Tenets
- “Maintain consistent brand styling.”
- “Accessibility is non-negotiable.”
Product Requirements Document
- Features
- User Stories
- Acceptance Criteria
- Low Fidelity Wireframes
Engineering Requirements Document
- Architecture & Data Flow
- API Endpoints
- Database Schema
- Performance & Scalability
- Testing & QA
- Deployment & DevOps
- Technical SEO/AIO)
Chronicle the tasks or features that have been completed, plus any relevant notes or decisions made along the way.
It’s a running Changelog (tracking built features & key decisions).
Living Backlog (next steps & new ideas)
The goal? Bridge the gap between planning & dev—so AI coding assistants don’t just generate code in a vacuum but actually follow a structured dev process.
But even with all this structure, AI still struggles with consistency. I’m now figuring out how to handle cases where AI needs to transpile code & follow prompts reliably.
As you know, coding assistants have limited contextual awareness, meaning they often:
❌ Miss dependencies
❌ Misinterpret instructions
❌ Lose track of previous logic
💡 I’d love to hear from others trying this approach!
- How do you get AI to transpile code reliably (e.g., Sudolang) across different languages or frameworks?
- What’s the best way to keep prompts structured so AI maintains consistency?
- Have you found an effective workflow to extend AI’s contextual memory for larger projects?
I’m just trying to figure this out too—curious to hear what’s worked (or hasn’t) for you.
