The four you'll meet most

Once you prompt regularly, the same handful of failures recur across every tool. Recognizing them on sight — and knowing the standard fix — is what separates fluent users from frustrated ones.

1. Confident hallucination

The model states something false as if it were certain — a fake citation, an invented statistic, a function that doesn't exist. It happens because the model is optimized for plausible text, and a confident wrong answer often looks more plausible than "I don't know."

Fixes: give it the source material instead of trusting its memory; ask it to cite or quote; add "if you're not sure, say so"; and verify anything that matters.

2. Ignoring instructions

You said "under 100 words" and got 300; you said "no preamble" and got "Certainly! Here is…". Long or buried instructions get diluted.

Fixes: move the critical constraint to the end, make it emphatic ("IMPORTANT: reply in one sentence"), and split a giant prompt into smaller asks.

3. Format drift

The first answer is clean JSON; the fifth has a stray comment that breaks your parser.

Fixes: show an exact example of the format, say "return only the JSON," and constrain the allowed values. For anything a program consumes, validate the output and re-ask on failure rather than hoping.

4. Bland, generic output

Technically correct, totally forgettable. Usually a sign the prompt was generic, so the model defaulted to the average of everything it has seen.

Fixes: add a role, an audience, and concrete constraints; supply an example of the voice you want; ask for a specific angle instead of "write about X."

Don't ask "why is the AI bad at this?" Ask "which of the four failures is this, and what's the standard fix?" Almost every disappointing answer is one of these wearing a different coat.

Try this: Next time you get a weak answer, label it — hallucination, ignored instruction, format drift, or generic — and apply only that fix. Keep a tally for a week and you'll see which failure you trigger most, and which single habit will help you most.