I've been using AI coding tools seriously for about six months now. Not dabbling — building real features, shipping production code, making them part of my daily workflow. Here's what actually happened.
Month 1–2: The Honeymoon
I went deep with Codex at the start. The speed was intoxicating. Scaffolding that used to take an hour happened in minutes. I was generating boilerplate, writing tests, and refactoring code faster than I ever had. The output quality was good enough that I started trusting it quickly — maybe too quickly.
Everything felt like a shortcut. I was shipping more, touching more of the codebase, moving faster. It felt like I'd unlocked a cheat code.
Month 3–4: The Frustrations
The cracks showed up gradually. Subtle bugs that passed initial review but broke in production. Patterns that looked clean but didn't fit our architecture. Code that was technically correct but didn't match our conventions, and the inconsistencies started compounding.
The bigger frustration was context. I'd spend time carefully framing a prompt, and the model would lose the thread on multi-step tasks. I'd get confident output that missed a constraint I'd mentioned three messages earlier. The overhead of checking and correcting started eating into the time I was saving.
Month 5–6: Recalibration
I stopped treating AI tools as a replacement for thinking and started treating them as an amplifier. I got more intentional about which tasks I delegated and which I kept manual. I developed better patterns for feeding context and verifying output.
I also switched primary tools. Claude gained capabilities around brainstorming, writing implementation plans, and executing in worktrees that changed how I approached complex work. Then Codex jumped ahead for surgical, focused implementations. The tools have different strengths, and I learned to match the tool to the task instead of defaulting to one.
Where I Am Now
AI coding tools are a permanent part of my workflow, but with guardrails I didn't have in month one. I use them more selectively, verify more carefully, and invest more in context management than raw prompting.
If I could tell someone starting today one thing: the tool is as good as your judgment about when and how to use it. The engineers who get the most from AI aren't the ones who use it for everything — they're the ones who know exactly where it fits.
I imagine I'll keep switching between Claude and Codex as the models evolve. That's not indecision — it's the rational response to a landscape that's moving this fast.