Common prompt mistakes and how to fix them
Five patterns that reliably give a worse answer than you meant, each with a one-line fix. Use it as a checklist when a prompt keeps returning something close but not right.
These five turn up constantly, across models, and each fix is one added sentence.
Asking for problems without permission to find none
Review this and list the issues returns issues whether or not any exist, because the model completes the pattern you gave it. Fix: add if there are none, say so in one line.
No output shape
Tell me about this gets a medium-length paragraph aimed at nobody. Fix: name the format, whether a list, a table, a fixed set of headings or a word limit.
Burying the instruction under context
Three paragraphs of background and the real question in the last sentence works less well than the reverse, because the model weights recent instructions more heavily. Fix: state the task first and put the supporting context and input after it in a marked block.
Asking for a rewrite without saying what must survive
Make this better, with no limit on what may change, lets the model change anything, including parts that were right: a spec number, a name, a fact. Fix: name what must stay exactly as written.
Vague comparatives
Better, cleaner and more professional mean different things to different people, and the model has to guess which you mean. Fix: swap the adjective for the concrete thing you want, such as shorter, a more formal register, fewer adjectives, or matches this example.
What sits under all five
Each fix is a constraint that takes a decision away from the model. In practice that is most of what better prompting means.
Notes from the author
Keep this open beside a prompt that is not working and test it against the five mistakes before you rewrite from scratch. Usually one added sentence fixes it.
For information only, not professional advice. You use it at your own risk and we are not liable for any loss or damage. See the disclaimer.
Tagged