DEV Community

sreehari bandi
sreehari bandi

Posted on

I've Been Thinking About a Better Way to Work with AI Coding Assistants

#ai #coding #productivity #softwaredevelopment

I've been experimenting with AI coding assistants quite a bit recently, especially for development work involving multiple repositories.

One thing I've noticed is that generating code is becoming easier. But managing the overall development process is still not that simple.

For example, consider a feature that requires changes in both the frontend and backend.

We might have:

  • Changes in the UI repository
  • New APIs in the backend
  • Changes to shared models
  • Database modifications
  • Testing across multiple services

We can ask an AI coding assistant to work on each part. But how do we make sure it understands the complete requirement? How do we track dependencies? And how do we ensure that the implementation follows the intended design?

I feel this is where having a structured specification becomes important.

Before asking an AI agent to write code, I'd like to have a clear understanding of:

  • What exactly needs to be built?
  • Which repositories are involved?
  • What changes are required in each repository?
  • What are the dependencies?
  • How will we validate the implementation?

I've started exploring this approach by building a small developer tool and experimenting with Claude Code.

I'm particularly interested in three things:

  1. Managing multiple repositories as a single engineering workspace.
  2. Generating reusable agent instructions from technical specifications.
  3. Allowing developers to choose their preferred AI coding agent instead of being locked into one provider.

I don't think AI should replace the engineering process. I think we need to improve the way we work with AI.

I'm still exploring the right architecture and implementation approach.

I'd be interested to hear from other developers:

How are you managing AI-assisted development when a feature spans multiple repositories? Are you using structured specifications, or do you mostly work through prompts?

I'm building this project in the open and would love to hear your thoughts, suggestions and feedback.

GitHub: https://github.com/srihari1247/spec-desk

If you're interested in the idea, feel free to explore the repository, raise issues or contribute.

Thanks!

Top comments (0)