What are the common ways prompting goes wrong?

When a prompt does not give you what you wanted, it is rarely a mystery. The trouble almost always falls into one of a few patterns. Knowing the pattern is most of the fix, because each one has a known cause and a direct response.

The four covered in this topic are: the model stating something false with confidence (hallucination), the model wandering away from what you told it to do (instruction drift), the model getting lost in an input that is too long, and the model guessing because your request left room to guess (ambiguity).

None of these mean the model is broken. They are normal behaviors of a system that predicts likely text rather than looks up verified facts. Once you see a failure as one of these patterns, you can reach for the matching fix instead of rewriting the whole prompt and hoping.

Three of the four share a root cause that points straight at the fix: drift, ambiguity, and getting lost in a long input are all the model filling a gap you left, whether that gap is a vague instruction, a buried rule, or a fact stranded in the middle of too much text. The fix in each case is some version of closing that gap yourself, being specific, stating the positive, trimming what you paste, placing what matters at the edges. Hallucination is different in kind. It is not filling a gap you left; it is the model producing a confident, wrong statement about something it was never given and does not actually know, a habit built into how these systems are trained and scored rather than a symptom of how you phrased the request. That distinction is why hallucination gets its own fix (permission to say “I don’t know”, grounding in sources) instead of the specificity-and-structure fix that handles the other three.

The pages that follow take each pattern in turn: why it happens, how to spot it, and what to change.