When Claude Code Gaslights You (And How to Break the Loop)

Here's the pattern. You hit a bug. You tell Claude Code. It writes a fix. The fix doesn't work. You tell it so. It apologizes and tries again. New fix doesn't work either. After 4 rounds, it starts telling you things like "actually, the issue might be in your Node version" or "the previous implementation was already correct, maybe try clearing your cache."

This is a confident-wrong loop and you have to break out of it yourself. The AI will not.

The Sign You're In the Loop

You're in the loop when any of these happen:

• The AI suggests something you've already tried (and told it didn't work)
• The AI blames the environment without evidence
• The AI says "the previous version was correct" about code you know was broken
• Each fix introduces a new regression that didn't exist before
• The diffs get bigger and more speculative

The worst one is "it's working for me" — the AI has no local machine. It cannot know whether it's working. If it says this, it's a pattern in the training data, not a fact about your code.

Stop and Reset

Step one when you notice the loop: stop adding context. Every additional "that didn't work" message makes the AI dig deeper into its wrong model. Reset the chat. Reset the branch to the last known good commit. Start over with the original bug description, but this time with more specifics:

• The exact error message, full stack trace, no trimming
• The inputs that reproduce it
• The expected vs actual behavior in one sentence each
• What you've already verified (the env is fine, the dependencies are installed)

Narrow the Scope

The confident-wrong loop gets worse with scope. If you're asking "fix my feature", the AI has 5 files to speculate about. Narrow it: "in this specific function, this line returns undefined when it should return an array — why?" Small scope forces the AI to actually look at the code instead of confabulating.

Ask for Uncertainty

Explicitly prompt for "what could cause this?" instead of "fix this". Asking for a fix gets you a fix-shaped output, even if the AI has no idea. Asking for hypotheses gets you a list of candidates you can actually investigate.

My favorite prompt when stuck: "List the top 3 things that could cause this bug, ranked by likelihood. For each, describe how I could verify or rule it out in 30 seconds." This reframes the conversation from "fix it" (wrong confidence) to "investigate" (calibrated uncertainty).

Write It Yourself

Sometimes the right answer is: close the AI, open the file, and write the fix by hand. Some bugs need human pattern matching. A race condition that only shows up under load, a CSS bug from specificity interaction, a build tool config that contradicts another config — these are all things where the AI is guessing and a human who looks at it for 3 minutes will see it.

Knowing when to drop out of the AI loop is a skill. The signal is: after 3 rounds of failed fixes, your probability of success with more AI rounds is lower than the probability of success with 10 minutes of focused human debugging. Drop out.

The Vibe Code Fix checklist has items under Hallucinations that specifically target this pattern — "API matches the version you actually installed" catches the AI referencing wrong API versions, which is a common cause of the confident-wrong loop.