Five prompt mistakes that make AI answers worse
Vague tasks, missing context, no format, stacked requests and starting over too soon. How to spot each mistake and the one-line fix.
, 2 min read
Most disappointing AI answers trace back to a handful of habits. None of them are hard to fix once you can see them.
1. A task you cannot check
"Help me with my presentation" gives the model nothing to aim at. Help how? Shorter, clearer, more persuasive, better slides?
Fix: use one verb whose result you could check. "Cut these ten slides to six, keeping the three numbers on slide four."
2. Leaving out what only you know
The model cannot see your audience, your goal or what you have already tried. Without them, it gives you the answer that would suit the average person in your position.
Fix: add one sentence of context for each of: who it is for, what it is for, and what must not change. That is usually enough.
3. Not saying what shape you want
If you need a table and get five paragraphs, you have to ask again. If you need two sentences and get a page, you have to cut it yourself.
Fix: name the format and the length. "A table with columns for option, cost and risk." "Under 80 words." "Bullet points, conclusion first."
4. Stacking five jobs in one message
"Summarise this report, find the risks, draft an email to the team and suggest a title" will get you four half-done jobs. The model splits its attention, and the later tasks get the least.
Fix: do them one at a time, or ask for a plan first and then work through it. Each step gets the model's full attention, and you can correct course between them. Our glossary entry on prompt chaining explains the pattern.
5. Starting again instead of steering
When an answer misses, many people delete the chat and rewrite the prompt from scratch. That throws away everything the model got right.
Fix: reply with a short correction. "Good. Half the length. Drop the jargon in paragraph two." Models are very good at targeted edits. If you find yourself making the same correction every time, move it into the original prompt.
A bonus: mixing instructions with material
If you paste a long email straight after your instructions with nothing in between, the model can confuse the two, especially if the email itself contains requests. Put your instructions first, then a clear marker like "Email:" or a line of """, then the material. See delimiters.
Try it on your next prompt
Take the last prompt that disappointed you and run it through the prompt improver. It checks for each of these mistakes and hands back a stronger version. For the full pattern, read how to write a prompt that gets a useful answer.