I started off with the intention of writing my new Blender addon by hand. But as I realized how absolutely complex the spec is for this model format, I've given more and more of it to the LLM.
I honestly don't believe I would ever have gotten anywhere with this thing if I didn't. It would have...
/https://i.s3.glbimg.com/v1/AUTH_59edd422c0c84a879bd37670ae4f538a/internal_photos/bs/2026/s/r/cL9DouRjmx9qxB2rAFMg/whatsapp-image-2026-07-12-at-19.38.29.jpeg)


Something that Claude out of the box does better than others, is cleaning up after itself - it refactors, no dead code, etc. I had to prompt that behaviour in my agent especially.
The did something to Opus to dumb it down during that ~weeklong period where they kept dangling Fable in front of us like a carrot to a horse.
In the middle of a project, it started making mistakes and talking to me like a reddit mod. I wasted a lot of time on a problem that was caused by Opus. Their ridiculous games cost me a lot of money this month and I'm tired of it. Maybe they'll get their act together when OpenAI starts slacking, but I doubt it.
I'll most likely be moving to Kimi or GLM if they don't drop the guardrails. I'm only paying for these things because they speed up my work. If they refuse to do the work, why would I give them my money?
Well, they're really the best, it's not that kimi does what claude does. I did a lot of hopping and every claude killer was a mistake. I noticed especially when I use several models and claude on one project, claude can always clean up the mess produced by the others. Kimi for example produced mountains of dead code for me.