Finding the “Sweet Spot”: Why I’m Moving Beyond Claude and Embracing Codex

In my last few posts, I talked a lot about my 3-pane, 3-worktree setup for parallel AI sessions. It felt like the ultimate productivity engine. But as anyone working in this space knows, the “perfect” workflow is a moving target.

After some deep debugging sessions and a bit of inspiration from Peter Steinberger’s recent deep dive into agentic engineering, I’ve had to rethink two major parts of my day-to-day: how many “brains” I can actually handle at once, and which brain is actually the smartest for iOS development.

1. The “Sweet Spot”: Why Two Sessions Beat Three

In my earlier setup, I was pushing for three parallel Claude sessions. On paper, more is better. In practice, I found myself hitting a cognitive wall.

When you have three complex architectural refactors happening simultaneously, the mental overhead of switching between them starts to eat into the time you’re actually saving. You spend more time “syncing” with the AI’s state than you do making decisions.

My current sweet spot is 2 concurrent sessions.

Two sessions allow for a perfect rhythm: one session can be grinding through a long build-test-fix loop while you are actively planning or reviewing in the second one. It keeps the momentum alive without the “chaos” of a third context-switch. It’s less about maximizing the machine’s CPU and more about protecting your own.

2. When Claude Hits a Wall: The SwiftUI “View ID” Nightmare

I’ve been a “Claude-first” developer for a long time. Its reasoning is usually top-tier. But recently, I hit a problem that Claude simply couldn’t solve: a massive SwiftUI refactor involving a subtle issue with the view reload system and the Identifiable protocol.

If you’ve done deep SwiftUI work, you know the pain—views not updating when they should, or re-initializing entirely because of a hidden identity mismatch. I tried every prompting technique in the book. I gave Claude context, file references, and “thinking time.” It kept spinning in circles, fixing one side-effect only to break another. It felt clunky. It felt like it was doing exactly what I said, but not what I meant.

3. The Codex Revelation: Inference vs. Instruction

After reading Peter’s post about just “talking to the code,” I decided to try the Codex agent (now a first-class citizen in the latest Xcode intelligence updates).

The difference was immediate. Where Claude was literal, Codex was intuitive.

I gave it the same SwiftUI problem—a huge, messy refactor—and almost no explanation. I didn’t have to guide it through the View hierarchy or explain the reload logic. Codex just got it. It inferred the best solution by looking at the architecture, fixed the reloading issue on the first try.

4. Why This Matters for iOS Development

For iOS and SwiftUI specifically, this is the “Vibe Shift” that actually matters.

• Claude is an amazing instruction-follower. If you know exactly what needs to change, it will write the code beautifully.

• Codex feels like an intention-inferrer. It understands the nuances of the Apple ecosystem and the “SwiftUI way” of doing things without being told.

In the world of complex mobile architecture, having a tool that “infers what you wanted to say” is worth more than ten tools that only do what you tell them.

Closing Thoughts

My professional reality is still defined by company policy, which keeps me tied to Claude for my daily work. To make that work, I continue to lean on strict planning phases and disciplined requirement docs to keep the model from hitting a wall. It’s a reliable process, but it feels like “manual steering” compared to what I’ve discovered elsewhere.

On my personal projects, I’ve moved to Codex, and the shift is night and day. I’ve stopped planning and started implementing. Where Claude requires a roadmap, Codex infers the intent, solving complex SwiftUI logic that Claude simply can’t grasp. Scaling down to two sessions with a smarter engine has proven that raw speed comes from removing friction, not adding more AI.

Takeaway:

• 2 Sessions > 3: Scaling down protects your cognitive load and actually increases your daily throughput.

• Policy vs. Preference: Claude remains my plan-heavy partner for work, but Codex is my choice for skipping the overhead and moving straight to code.

• The Sweet Spot: Productivity isn't about juggling more sessions; if the model understands your intent straight away, you don't need the extra noise to move faster.