A few weeks ago, I switched from ChatGPT to Claude for day-to-day LLM use. I’d been thinking about it for a while. Claude has been shipping features aggressively. Someone on Threads counted over 50 new features in just a few months. It gets good reviews from developers, marketers, founders. People actually using it for work.
I was happy with the switch. More than that. It changed how I approach tasks. It rewired my instincts.
Now, no matter how small the task, I want to build an app for it.
Planning YouTube content? Content planning app. Long-form writing? Outline the chapters, then build a writing app with a chat companion baked in. SEO audit? Chrome extension. Updating my LinkedIn profile? An HTML tool that spits out recommendations.
People are building dedicated apps for every little personal workflow. That’s not a trend anymore. It’s just how people work now. I’m all in.
Then reality hit.
Claude doesn’t handle the current volume of users well. Sometimes the agent skips a step, misses a skill file, or loses context and solves the wrong problem. I run Sonnet by default. I figured it was enough for most things. I haven’t even tried Opus. It costs five times more per token than Sonnet. But even tasks that should be well within a Pro account’s daily limits, debugging a WordPress theme, nothing exotic, can drain it.
I’m not alone. I’ve seen complaints about this on Threads, even from Max plan users on 5× or 20× higher limits.
Around the same time, I started seeing Codex promotions. Generous limits. Compelling pitch. I resisted because I like Claude.
Until today.
So I tried it.
I’d never used ChatGPT for actual coding. For me it was always a brainstorming tool — video ideas, narration, shot planning. Never for building something like a WordPress theme. But the token frustration, plus weeks of OpenAI noise, pushed me to give it a shot. I resubscribed to Plus.
I started by figuring out how to use both tools without having to choose. My first instinct was to build a new infrastructure from scratch. Then I realized I already had one with Claude. Not perfect, but solid, built step by step as I learned.
The thing is: those same preference files could work for both.
So that’s what I did today. I aligned Codex and Claude around a shared setup. I went back and forth between them, iterating. Eventually Codex finalized the preference files and put together a system map. Which task goes where, across which projects.
After a few tests: it works for both.
Over the next few days I’ll keep refining it. For now, I’m splitting the work by feel:
Claude handles planning and writing code. Codex handles debugging.
I’ve also tried flipping it — having Codex debug and rewrite Claude’s output, then Claude validate before deployment. Still testing which split makes more sense.
What I have now: a local folder both agents pull from as a shared preference system. Portable across machines. Backed up with Git.
Three weeks ago I was a complete beginner with agentic AI.
Today I have a business built around it.
And two agents on the payroll.