General
Share and discuss tech, products, business, startups, or product recommendations
Start new thread

One thing AI coding tools still don't get right

by•

Over the past few months, I've noticed that AI coding tools are excellent at generating code, but they still struggle to keep context across larger projects. That often means spending time re-explaining the codebase instead of moving faster.

I'm curious whether others have noticed the same thing, or if you've run into different limitations. What's the biggest issue you've experienced?

62 views

Add a comment

Replies

Best

For me, AI still struggles with understanding long-term context. I've found that I spend more time explaining intent than writing code.

 One thing I've learned is that AI can generate solutions quickly, but it rarely understands why I made earlier design decisions. I still have to reconnect every new suggestion to the bigger picture, and that's where I spend most of my time

   For me, the hardest part isn't code generation, it's continuity. I've noticed AI often forgets subltle constraints that shaped previous decisions, so I end up repeating myself.

 One thing that stands out to me is how AI handles refactoring. I've seen it improve individual files, but it sometimes overlooks the ripple effects across the rest of the codebase.

I get you, bro. It's a daily struggle for vibe coders. A new session means a need to set new context and start over. Some things that I do that has helped me to keep better context
1. Set instructions that are around your work model
2 - Gather all your sessions in projects that has all the reference files, prd, architecture, research, etc.
3 - Always keep updating your or a PRD file and set project-specific instructions
4 - Think in AI chats and execute in CLI or code editors.

My biggest issue is that AI can still make up incorrect info while coding then

Yes, context is still the biggest weakness for me. AI is great at writing code, but once a project gets bigger it can lose sight of earlier decisions. I end up spending more time restoring context and reviewing changes than actually generating the code.

Just tell it to write plans and designs in markdown files and refer to those documents when building or modifying things. Then you don't have to remind it of anything or explain as much: the markdown files remain the source of truth.