> [Metrics] from the Vibe Kanban - a tool which orchestrates AI agents - has shown Claude Code usage drop from 83% to 70%
I'm not really convinced that this warrants the title the post currently has. For one, I hadn't even heard of Vibe Kanban prior to this, and for two, the error bars on this must be insanely wide as is.
For most of the vibe coding crowd the novelty simply wore off. You can only tinker with small projects for so long before craving something more substantial. When the tool inevitably struggles with your evolved, more complex goals, you perceive it as having gotten worse.
Codex just seems to have a much bigger context and doesn't chew up tokens as readily as Claude. It seems to be able to do a much wider and broader range of things, accepting much wider and broader instructions and implementing them perfectly.Whereas Claude would struggle.
Poor performance, over tight rate limits, inability to follow prompts suddenly, erratic outputs, which together have combined to increase project costs massively?
a dev in philipine india and pak will turn out to be cheaper and better than ai
Junie is cheaper and built into Webstorm. Also Claude tried to charge my card for months after I've tried to cancel it. Despicable.
Among the worst things that I have a hard time tolerating about Claude:
- sycophancy, I'm honestly tired of "You're absolutely right". I need a pair programmer, something that's gonna correct me, provide different ideas, etc.
- inability to follow the script. Even though it will tell you you're right, it will still do its thing. Doesn't matter if I spend 2 hours writing a detailed spec file, a todo list, etc, Claude's gonna do its thing regardless. You can even correct it with "no, don't do this", and it will still do it regardless. I understand that this is how AI works (it's like children, if you tell them to not do something it's more likely they will), but it's just unbearable.
For both of these things it's impossible to make it go right. No matter the system instructions, the prompts, the context management, it's just terrible.
That's not to say it's all bad: there are things I like about Claude and AI assistants. I firmly believe that a coder with AI is much more productive than one without. But what AI should be delegated to, is not writing and editing code, but planning it, writing specs, doing research, verifying you're maintaining docs, suggest ideas, alternatives, test cases, reviewing PRs according to guidelines, etc.
I don't even think it's a matter of "it will get better", it produces way too much code than a human can review and reviewing code is more difficult than writing it in the first place.
Even more, it can provide its value in tasks humans are just bad at such as writing good issues/tasks, stuff like user stories that use consistent ubiquitous language, etc, etc. Stuff that it's hard to get stakeholders to get right, but can be achieve with a set of good rules and having the stakeholders interact with the chatbot first that can guide them writing much more clearly.