With AI design iteration, never use the word 'mock-up.' Even if you say 'design a mock-up as if ready for production,' it still skews toward a lesser design than it's capable of.
2 weeks ago hit my Claude weekly limits 2 days early
had an actual panic attack.
At the time I didn’t trust Codex with my code.
Since then I got a $100 subscription to Codex
Been using it as my main driver over Claude.
Feels so ruthless. I have lingering feelings about it
3 days. Opus 4.7 and GPT-5.5 couldn't crack it.
Gave the same problem to Kimi K2.6 and DeepSeek v4 solved on the first try.
Now I run Opus, Kimi, and DeepSeek in parallel and let GPT-5.5 referee.