Context mistakes
Giving the agent too much or too little to work with.
Too little context, or too much
Most Cursor problems are context problems, and they come in two shapes. Too little: no rules, a stale index, or a prompt that never points at the relevant files, so the agent guesses and drifts. Too much: dumping half the repo into a request so the signal is buried and the model loses the thread. Both produce confident, off-target edits, and both are fixed the same way — by being deliberate about what the AI sees. The model is rarely the problem; the context around it usually is.
Aim for the relevant slice
The fix is to give Cursor the right slice: the files the change touches, the conventions it should follow, and a scoped task it can complete and you can review. Set up rules so your conventions are always present, keep the index fresh so it can find related code, and use @-mentions to point at the key files. Then keep each change small enough to verify. Get the context right and the edits improve immediately, without changing anything else about how you prompt.
Why does Cursor make off-target edits?
Almost always a context problem — too little (no rules, stale index, no file pointers) or too much (half the repo dumped in). Both make the model guess or lose the thread.
How do I fix bad context in Cursor?
Give it the relevant slice: set up rules, keep the index fresh, @-mention the key files, and scope each task. Deliberate context turns off-target edits into accurate ones.
Can too much context hurt in Cursor?
Yes. Dumping too much into a request buries the signal and makes the model lose focus, just as too little starves it. The goal is the relevant slice, not everything.