Range vs Credit Anxiety. Anyone? 🤪

If you have read my previous posts, I have shared how Claude Code performance tanks once your code and best practices hit a certain size due to context bloat.

(Link: https://lnkd.in/gTwiKkz3)

And the next natural thing to try is to break your best practices into skills and use specialized agents to enforce them at every step, from plan, to plan review and code review.

It has worked pretty well after I tested it.

Hallucinations drop and reliability boosts.

But now, there is a new problem. It burns at least 2x more tokens.

So, what is next?

Maybe it is time to test local LLM?

It would be interesting to know if more loops with local LLM can get the same performance as Opus 4.5.

This would be my next experiment. Stay tuned.

Have you tried running local LLMs for coding tasks? How does the quality compare to cloud models for you?

#AIEngineering #VibeCoding #LLM #ClaudeCode #SoftwareEngineering