The weak link #1

Dropped in to say cheers for this project.

Read your article and the approach is very cool! Good job!
I used it on a typescript features library and I like it.

However I want to mention that using it with Claude Code, I caught him lying and ditching the DevFlow protocol.

It seems that the protocol is left behind when I start asking questions about future decisions. Claude seems to take over. He seems to fake/replicate the behavior of DevFlow but it's just a trick.
I asked him if he ditched the protocol and he said 'Yes, even if I told him to follow it specifically'.

So I think the weakness is in the LLM itself as the orchestrator seems to work indeed!

What I found is that is better to open a new conversation for each new task.

Cheers