Taran RampersadRE: https://mstdn.social/@hkrn/116472030680794257
Agentic AI didn't do a whoopsie.
It chose to do things it should not have. 🙄
And a bad architecture choice... blammo.
Pretty sure lawyers are looking over fine print right now.
Wulfy—Speaker to the machinesIdiots should not #vibecode
I don't know what guardrails cursor has, as Claude code is my thing.
In my workflow, I have claude.md which contains operational directives including backups and docs.
NO COMMAND GETS EXECUTED WITHOUT MY AUTHORITY.
My vibecoding sessions consist (after issuing the prompt) of 1-40 consent commands for every single command that Ai executes.
1. Don't do it
2. Execute blah blah
3. Stop (where I can amend/seek explanation then proceed)
At every step Claude explains what it is doing, I constructed my .md this way.
Sure it may burn more tokens, but I have an idea of the process.
The code monkeys got lazy.
Lesson learned I expect.