Claude Code confirms edits but changes don't reflect in running app

[SaadHadadia]

I FUCKING HATE VIBE CODING!

I was hired to help out on an ongoing project in a technology that I am not familiar with , so I decided to vibe code my way through it with Claude Code.

Initially, it was working great. However, somewhere down the line, something must have gone wrong. I have found that when I ask it to implement a new functionality or fix an error, it assures me that all modifications have been made successfully, yet when I test the application myself, there is no visible difference at all. TOKENS ARE BURNING AND YET THERE IS NO PROGRESS AT ALL!

Moreover, the codebase has become such a mess that debugging manually becomes almost impossible.

Has anyone else ever experienced this issue? How did you deal with it?

[affaan-m]

That failure mode is real. It usually comes from one of two things: edits are landing in a different tree than the running app, or the app is no longer in a state where small edits produce observable behavior.

Before spending more tokens, make the agent prove the loop:

1. Run pwd, git status --short --branch, and git diff --stat before/after the requested change.
2. Confirm the dev server/container is running from that exact checkout.
3. Restart the dev server once to rule out stale hot reload or build output.
4. Ask for a tiny visible sentinel change in the exact component/route you are testing, then remove it after verification.
5. If the codebase is already messy, stop asking for broad feature work and switch to one failing test/reproduction, one small diff, one verification command.

ECC cannot make Claude Code magically observe your running browser, but it is designed around that discipline: inspectable diffs, explicit checks, and no claims of success without evidence. If you share the stack, how the app is launched, and whether Docker/WSL/hot reload is involved, people can suggest a tighter debug loop.