How has shipping changed for me? The gap between idea and working code has collapsed. What felt like magic in May—prompts that actually worked—is now baseline. I ship at a pace that still surprises me.
The “vibe coding” discourse misses the point. Spend enough hours with agents and you develop intuition for how long things should take. When Codex returns empty-handed, I get suspicious before I’ve read a line.
GPT-5 was the unlock. It took weeks to unlearn Claude patterns and trust the model’s appetite for context—those long silent reads where it ingests files for fifteen minutes before touching code. Annoying, until you realize it fixes the right thing. Opus rushes. Codex studies. The benchmarks lie; you feel the difference in rework.
Most software is plumbing. Data in, data out, maybe store it, maybe show it. I default to CLI—agents can call it, verify it, close the loop. Hard thinking is the bottleneck now, and honestly, most projects don’t need much.
I run 3-8 projects simultaneously. Context-switching is expensive, so I batch it for deep-focus days. One anchor project, satellites orbiting. The queue fills with ideas; Codex churns through them while I write, eat, sleep. No elaborate orchestration—I’m usually the bottleneck anyway.
I don’t branch. Don’t checkpoint. Don’t revisit old sessions. Linear history, docs folder for context, cross-reference solved problems from other repos. When something’s ugly, I prompt the fix. Building software is circling a mountain, not climbing straight up.
Dependencies and system design still require human judgment. Everything else is inference-limited.