The Prompt Mistakes Everyone Makes at First

The short answer

The most common prompt mistakes are vague requests, missing context, cramming everything into one giant question, accepting the first answer, asking the model to vouch for its own accuracy, and reusing one favourite prompt for every job. Nearly everyone makes the same handful when they start, which is good news: they are all fixable, and none of the fixes takes longer than an afternoon to learn.

None of these mistakes means you are bad at this. They happen because talking to an AI assistant feels like talking to a person, and a person would fill the gaps with common sense and knowledge of your situation. The model has neither. It only has your words. Every mistake below is really the same mistake wearing a different hat: leaving a decision to the model that only you could make.

Mistake 1: the vague ask

"Make this better." "Write something about the project." "Can you improve my CV?" The model cannot ask what "better" means, so it guesses, and the guess is always the most generic possible reading of your request.

The fix is to say what better means to you. Shorter or longer? Warmer or more formal? For whom? Here is the difference on a request people type every day.

Mistake 2: missing context

The model does not know your company, your customer, what happened last week, or what you have already tried. Beginners leave all of that out because it feels obvious, and it is, to them. The model then invents a plausible version of your situation and writes for that instead.

The fix is a habit, not a technique: before sending, ask yourself what a smart stranger would need to ask you before doing this task, and answer those questions in the prompt. Names, dates, audience, background, constraints. The details that feel too obvious to mention are usually the ones doing the most work.

Mistake 3: one giant question

"Research this, summarise it, draft the report, make it persuasive, and format it for the board" is not one task, it is five. Packed into one prompt, each task gets a fraction of the model's attention and you get no chance to steer between them, so one early wrong assumption poisons everything after it.

The fix is to break the job into steps and feed each answer into the next. That technique has a name, prompt chaining, and it deserves its own explanation: What Is Prompt Chaining? walks through a full worked example. The short version: ask for step one, check it, correct it, then ask for step two.

Mistake 4: accepting the first answer

Beginners treat the first output as the verdict. If it is bad, they either give up or delete everything and start again. Experienced users treat it as a first draft from a capable assistant who has not met them before, and they reply with feedback.

The fix is one round of specific criticism instead of a restart. Not "try again", which just rerolls the dice, but feedback that names what to keep and what to change.

Good structure, keep it. Now change three things:

[what is wrong and how to fix it, e.g. the tone is too formal, write it like a colleague]
[e.g. cut the second paragraph, it repeats the first]
[e.g. lead with the deadline, it is buried at the end]
Change nothing else.

Two or three rounds like this almost always beats starting over, and it is much faster than editing the whole thing by hand.

Mistake 5: asking the model to police its own accuracy

At some point every beginner discovers that models state false things confidently, and the instinctive response is to write "do not hallucinate" or to ask "are you sure?" and trust the reply. Neither works the way people hope. The model does not know which of its statements are false, so it cannot promise to avoid them, and it will answer "are you sure?" with the same confident fluency whether the underlying claim is true or not.

The fix is to change what you verify, not what you ask. Anything checkable that matters (a figure, a date, a name, a legal or medical claim, a quotation) gets confirmed against a real source before you rely on it. You can make that easier by telling the model to separate facts from suggestions, or to flag anything it is uncertain about, but the flag is a starting point for your checking, never a substitute for it.

Mistake 6: one mega-prompt for every job

Somewhere online, every beginner finds an enormous "ultimate prompt" full of role-play instructions and formatting rules, and starts pasting it in front of everything. The trouble is that instructions written for one job actively hurt another. A structure that suits a report smothers a quick brainstorm, and a tone spec for client emails is wrong for a note to your team.

The fix is to keep a small set of short prompts matched to the jobs you actually repeat, rather than one long prompt for everything. A useful skeleton to build them from:

Task: [one clear verb and object, e.g. summarise these notes]
For: [audience]
Context: [the facts only you know]
Format: [bullets, table, word limit, whatever fits this job]
Avoid: [the failure you keep seeing on this kind of task]

Five lines, adapted per task, beats five hundred words of borrowed boilerplate every time.

How to practise

Pick one real task from your week, something you would do anyway, and run it twice: once with your normal prompt, once after fixing whichever mistake above you recognised in it. Comparing the two outputs teaches more than any article, because the improvement is on your own work.

If you want the structured version of that practice, our free Prompt Engineering course covers each of these habits properly, with interactive examples and no account needed. And if you would rather start from prompts that already work, 11 AI Prompts That Actually Save You Time at Work gives you eleven to copy today.