Testing a New AI-Integrated Dev Workflow at Work

At work, I recently proposed a new workflow to better integrate AI tools into our development process. Here’s the outline of what we came up with:

1. The PM writes a PRD (Product Requirements Document) in Markdown on GitHub.

2. Developers and stakeholders review the PRD using GitHub’s review features.

3. Once finalized, developers generate detailed project specs, including design docs for each feature.

4. Test cases are written with acceptance criteria baked directly into the tests.

5. During development, engineers feed these documents into their AI coding assistants (I use Cursor).

6. If we rely on information outside the codebase, we may need to create MCPs. Creating MCPs could eventually become part of our official responsibilities.

7. Developers audit the AI’s output.

8. QA is conducted.

9. Release.

We’re currently piloting this workflow with a small new feature. The goal is to evaluate how well it works, especially in terms of AI effectiveness and development efficiency.

Key questions we’re trying to answer during this trial:

• What’s the optimal task size for an AI assistant to handle without generating nonsense?
• What additional context can we feed the AI to reduce hallucinations?
• How much time can this workflow actually save?