thorny gust
I have been testing Kiro now for two weeks as a pure vibe coding tool and for this I would give it a fail. However, I do believe Kiro is on the right track to pass this and I hope below offers good constructive feedback.
A bit of background: I am the Head of Engineering at my company and the interest in this tool was to expand who can work on the code, given sufficient guidance. I subbed to the $20 and I am now on my way to 1000 vibes overage - which I'm hoping is going to be reimbursed as promised!
Anyway, the feedback (happy to chat on these) in no particular order:
-
Kiro doesn't work well with Windows terminal. It tries to run scripts and commands that don't give feedback unless I ctrl-C. A simple one is doing npm run dev - if it runs the Kiro is stuck forever.
-
It does not have memory of conversations, ideas, improvement. I'm bypassing this by asking it to update documentation every time but this should be automatic. Even better would be contextual RAG memory.
** This is especially important when working on multiple features and an improvement on one impacts another. -
Stop saying I'm absolutely right.
-
Tasks are marked as complete are 90-95% of the time not production ready (not exaggerating).
-
For every 1 vibe I spend about 5-10 vibes fixing code.
-
Clean up after yourself! It creates a million test scripts and leaves it there.
-
Does not listen to steering documents (Kiro even admits this).
-
I spent a huge amount of time trying to get good specification documents - Kiro simply misses important details. I then spent that time in other AI products to really get exactly what I wanted. I then fed this to Kiro to generate specs, design, requirements and again, it oversimplifies - missing the core business logic.
-
Not listening to instructions and then lying about having done it.
-
Breaking featureA in order to make featureB work.
-
Creating an enhancedFeatureA instead of refactoring featureA. We now have two different "featureA" code.
paper nacelle