Nine prompt mistakes that ruin the answer you get back

If the answer you got back was vague, generic or plain wrong, the cause is usually sitting in the question rather than in the model. A chatbot cannot ask you the five clarifying things a colleague would ask before starting. It fills the gaps with the most statistically ordinary version of your request, which is exactly what "generic" means. Below are nine mistakes, grouped by what they do to the answer, each with the symptom you would recognize, why it happens, and the shortest fix. Work down the list when something comes back wrong and you will normally find your problem in the first three.

Mistakes that starve the model of information

1. No context about who you are or who it is for

Symptom: the answer is technically correct and useless. Advice pitched at nobody in particular, or an email that sounds like a press release.

Cause: you know that this is for a 40 person dental practice in Ohio and the recipient is an annoyed supplier. The model knows none of it.

Fix: add two sentences of situation before the request. Who you are, who reads it, what has already happened. When the same situation comes up daily, custom instructions that sit in front of every message can carry it for you.

2. Pasting a wall of text with no instruction

Symptom: you paste six pages of meeting transcript and get back a summary of the whole thing when you wanted the decisions.

Cause: with no instruction attached, summarizing is the default behavior, and it averages everything equally.

Fix: put the instruction first, then the material, and say what to ignore. "From the transcript below, list only decisions made and who owns each. Ignore scheduling talk."

3. Asking for sources without asking for verification

Symptom: five references that look perfect, two of which do not exist.

Cause: a model that is not connected to live search produces text shaped like a citation, because citations have a very predictable shape. That is how invented sources appear.

Fix: ask it to name only sources it can quote a line from, or turn on the tool that actually searches, and then open each one yourself. Checking that a title and an author exist together takes about 20 seconds per source, and is the first move in fact checking an AI answer properly.

Mistakes that confuse the instruction

4. Stacking several questions into one message

Symptom: question one gets four paragraphs, question three gets a sentence, question four vanishes.

Cause: the model is producing one continuous answer and allocating attention unevenly across a compound request.

Fix: one message, one job. If they truly belong together, number them and ask for the answer numbered to match, or run them as a chain of smaller prompts that feed each other.

5. Saying what you do not want instead of what you do

Symptom: "do not use jargon" produces a reply full of different jargon. "Do not mention price" and price shows up in paragraph two.

Cause: a negative rule defines an enormous space of acceptable answers and gives no target inside it. It also keeps the forbidden topic active in the conversation.

Fix: state the positive version. Not "do not be formal" but "write it the way you would text a colleague you like." Keep at most one or two negatives for genuine hard limits.

6. Asking for "brief" or "detailed"

Symptom: brief comes back at 600 words. Detailed comes back at 300.

Cause: those words have no fixed value. The model has seen them attached to every length there is.

Fix: use a countable unit. "Three bullets, maximum 15 words each." "Under 120 words." "Exactly five rows." Counting is not a model strength, because it works in tokens rather than words, so expect it to land close rather than exact, and check anything where the number really matters.

7. Leaving the format to chance

Symptom: you wanted a table you could paste into a spreadsheet and you got prose with a friendly preamble.

Cause: unspecified format defaults to conversational paragraphs, because that is what most training text looks like.

Fix: name the structure and show a one line template with placeholders. The mechanics are covered in getting output in the format you need.

Mistakes in how you handle the answer

8. Continuing in a chat that has drifted

Symptom: you corrected it twice and it keeps returning to the original wrong assumption.

Cause: every turn of the conversation is resent to the model each time, so the mistake, your correction and the model's apology all sit there together competing for influence. Long threads also push early instructions toward the edge of the context window, where they get less weight.

Fix: start a new chat. Paste in the good material and the corrected assumption, and drop the argument.

9. Treating the first reply as final

Symptom: you judge the tool on its opening answer and conclude it cannot do the job.

Cause: the first answer is a draft produced with the least information the model will ever have about what you want. Your reaction to it is the information it was missing.

Fix: respond with a specific edit. "Too abstract, give me the version with the actual numbers from the table." "Cut the intro, start at the recommendation." Two rounds of this beats an hour spent perfecting the opening prompt.

A fast diagnostic table

What to check first next time

Before you send a prompt that matters, read it back and ask one question: could a stranger with no knowledge of my situation produce the answer I have in my head from this text alone? If not, the missing piece is what you add. When the answer still disappoints after that, the positive version of this page is how to write a prompt that gets a usable answer, and if the problem is accuracy rather than shape, start with why AI invents things.