Two weeks ago, I spent far too long trying to build an NPM package called Clauded, which aim to use the hooks available to control the output from Claude code to include a confidence score on suggestions and approach.

The idea was simple enough and came out of the experience of looking at the outputs not feeling confident that they were good and asking the AI to reflect on its own confidence and seeing it tell me that only felt 20 or 30% confident that it suggestion would resolve a problem, which felt like something that should be more transparent unavailable to the user.

Ultimately, I couldn't really get the hooks to work reliably and a mixture of Cursor rules and Claude own interpretations of what I was asking you to help me write ended up influencing the output in a way that it would appear to be working and then as Claude's memory drained it worked in consistently and ultimately it wasn't really the script running but towards interpretation of how the script should run - it was really an echo chamber experience.

I've had a number of issues with the AI running Git commands in a way that's really destructive to my safety net when I'm working on a project which is committing those changes to a branch, so I've decided to have one more go at getting the AI to do what I want but this time it will be a set of controls around using and protecting Git from the AI.

Basic Set-up:
https://github.com/qstudio/alined-git-npm