Quirks are pretty much unavoidable. I tend to get better results using Codex. It sticks to established patterns. Slow, but more deliberate. Claude focuses more on speed.
A well-documented codebase lets both developers and agentic models locate relevant code easily. If you treat the model like a teammate, extra docs for LLMs are unnecessary. IMHO. In frontend work, code moves quickly.
I’m definitely interested in reducing token usage techniques. But with one session one problem I’ve never hit a context limit yet, especially when the problem is small and clearly defined using divide-and-conquer. Also, agentic models are improving at tool use and should require fewer tokens. I’ll take as many iterations as needed to ensure the code is correct.
Writing and updating CLAUDE.md or AGENTS.md feels like pointless to me. Humans are the real audience for documentation. The code changes too fast, and LLMs are stateless anyway.
What’s been working is just letting the LLM explore the relevant part of the code to acquire the context, defining the problem or feature, and asking for a couple of ways to tackle it. All in a one short prompt.
That usually gets me solid options to pick and build it out.
And always do, one session for one problem.
This is my lazy approach to getting useful help from an LLM.
Let cc/codex do research by themselves.
I usually ask claude to scan codebase about specific part for context, describe my issue or feature then ask it to suggest solutions.
Example :
Study <subject>, I want to <fix/add feature>. What do you suggest ?
From there, I have 2-3 different strategies options. Pick one, refine then implement.