State of the Slop

"This is a weird one. With this I went more hands-off by prompting my OpenClaw (which itself is a vibe-coded slop) running Qwen3.5 to call Codex with GPT-5.4 medium, and then switched to OpenCode with MiMo-V2-Pro when that ran out of quota.

When the switch happened, what the frontier GPT model left us was an absolute slop shit overengineered monorepo with a backend for account log-ins and "social features" and the worst stereotype-gradient-bluish-round-hover-border-slow-animations generated retard webapp feel you can think of with broken layouts and weird useless not-even-working pages and an inaccessible course itself (some fault in my initial prompting to make it a "platform"). Then MiMo turned it into the near-current minimalistic design with no extra features and a working course in... 1 turn.

I mostly just focused on higher-level prompts to the claw and it then prompts the coding CLIs with more specific actionable feature requests in the AI-fasioned markdown format. I didn't really expect it to work this well... the taped-together double prompting. I later checked on the instructions it gave to the CLIs and they were very accurate, detailed, and high-effort from an AI whom I thought would have little understanding with the codebase itself.

After some of the rounds, everything would seem to have turned into an unfixable mess with different truly unrelated things (for example broken dark mode, badly styled buttons, and exercise questions containing no stem) just broken or made no sense, where I thought would need intense hand-holding and explicit fixing rounds on all the broken parts I found. No, explicit prompting to fix one of those issues did nothing to fix that issue, and an unrelated iteration turn would somehow fix ALL of them and then everything is so back again. We truly don't understand modern technology.

Still some level of manual checking of the codebase going on, but less."