I don't know what I'm doing wrong to not have success with it the same way others do. I spent a few hours yesterday trying to get it to return correct texture size for one 3d model used in my Godot POC, that it generated itself in Blender. The model is a wall tile (so a simple 3D rectangle) and the texture is 1024x1024 and it struggled to return the specs. I wanted to commission an artist to paint one for me. Every time I went back it would correct itself:

- I told you from memory, not from the actual game itself - I gave you the texture size that we use today, not what we need for the game - I don't know why I said that, but here's the correct information - Sorry that wasn't actually correct, you're right to push back, here's the actual correct information - I didn't consider top of the wall that needs to be covered, sorry here's the latest correct information - Yes you're right, sides needs painting too, that's my mistake, here's what you actually need for the artist - No you don't need to paint the back side, it's never visible in the game - (me: prove it) Sorry you're right, back side is visible so you'll need another texture ... Back and forth, back and forth for eternity. I use Opus 5.5 at home and Fable at work and since I don't spend a lot of tokens at work due to mostly using it for operations, I use sometimes to help me with specific personal issues and it I see 0% difference between Fable and Opus.

If I ask a human person to tell me what is required to migrate a Key Vault in Azure from one tenant to another, they would elevate privileges, find the key vault, look up role assignments, access policies, secrets & keys and then tell me. Claude on the other hand forgets to look up role assignments and returns secrets but not certs. Then it goes on to migrate one, causes a clusterfuck, realises what it had done, tries to fix it and we go back and forth checking everything just wasting time and tokens.

This has been my experience for the last 2 years using it daily for work, nothing but frustration. I have instruction files in each working dir and hooks built-in on launch to read those and block it from saving anything to memory unless it's saved to local filesystem first and I'm happy wasting tokens for that but it still does not follow instruction. It will randomly bypass its hooks or on launch only read first 2kb of a file etc. It's not just frustrating, it's worrying as well because it can't explain why it's doing that.

I always get confused when I see these posts or testimonies on the internet from people who built Adobe clones with a single prompt and can hack NASA and invent new vaccines on a lazy Sunday afternoon. That's not my experience. Like at all. Anything that is not very clearly defined, it can't fill in for. I don't understand this. The entire advertised value of using AI isn't automation, it's autonomy. And I can't trust it to be autonomous because it causes more problems that I have to clean up than it's worth the time it saves on automation. I can't do the simplest of things unless I hold its hands and tell it in details what is it what I need from it and how to do it, and if that's true then I get very little value out of AI because at that point might as well just finish the task myself.

That's not how I work with people. If a new person joins our team, they will be handed a task and told to try and figure it out. When they get stuck I explain, rinse and repeat. Every time I explain a different thing, never the same thing. Because people learn and in a short time they can be autonomous with limited oversight. If Claude was a person it wouldn't pass probation.