AI Tools

10 Common AI Prompt Mistakes and How to Fix Them

Most bad replies are bad prompts. Each fix here takes under a minute.

Alina Tech & AI Writer

Sep 05, 2026 11 min read

Diagnose it from the reply

Start with what came back. The symptom points to the mistake.

What the reply looks like	Likely mistake	Fix
Generic, could be for anyone	Vague task or missing context	1, 2
A wall of text	No format or length	3
Half the jobs done, the rest muddled	Too many tasks in one prompt	4
Right facts, wrong style	No example	5
Still does the thing you banned	Only negative instructions	6
Invents details about your project	Context you never pasted	7
You keep opening new chats	Restarting instead of iterating	8
Confident numbers you cannot find	Facts taken on trust	9
Long prompt, no better reply	Over-prompting	10

Before you type Mistakes 1 to 3: the task itself

Ten seconds of thinking saves three rounds of replies.

1 - Vague task

“Help with” is not a task. The model picks one interpretation and you get the average of all of them.

Symptom: a polite, generic reply that answers a different question.

Before

Help me with my resume.

After

Rewrite the summary section of my resume for a data analyst role. Under 60 words. Resume below.

Fix: verb, object, scope. “Rewrite the summary, for this role, under 60 words.”

What is happening	“Help” fits rewriting, critiquing, formatting and reassuring, so the model hedges across all of them.
Why the fix works	The verb picks the action, the object picks the target, the scope sets the limit. Three words remove three guesses.
Applies when	Any prompt under ten words.

Anthropic's be clear and direct guidance shows the same fix with worked examples.

2 - Missing context

The model does not know who will read the output or why it exists. It guesses, and guesses safely.

Symptom: correct but flat; nothing in it is specific to your situation.

Before

Write a welcome email.

After

Write a welcome email for new users of a budgeting app who signed up from the free-trial page. Goal: get them to connect a bank account today.

Fix: say who it is for, why it exists, and what should happen next.

What is happening	Without a reader and a purpose, the model writes for the average reader, who is nobody.
Why the fix works	Audience sets the vocabulary, purpose sets the emphasis, the next step sets the call to action.
Applies when	Anything someone else will read.

3 - No format or length

Left to choose, the model picks long and structured. You then spend time cutting.

Symptom: headings, sub-bullets and 600 words when you wanted five lines.

Before

Explain SEO.

After

Explain SEO in 5 bullets, each under 15 words, for a bakery owner. No headings.

Fix: name the shape before the model picks one. Count, length, structure.

What is happening	Long, structured answers are the model's default because they are rarely marked wrong.
Why the fix works	A count and a length are hard constraints it can satisfy exactly; “concise” is not.
Applies when	Whenever the reply goes somewhere with limited space.

Inside the prompt Mistakes 4 to 7: what you put in it

What is in the prompt is all the model has.

4 - Five tasks in one prompt

Stacked instructions compete for attention. The first and last get done; the middle blurs.

Symptom: some tasks skipped, outputs merged, the translation half done.

Before

Summarize this, fix the errors, rewrite it, suggest a title and translate it to Spanish.

After

Step 1: list factual errors in the text below. Stop there. (Next turn: fix them. Then: title. Then: translate.)

Fix: one task per turn, or numbered steps with labelled outputs.

What is happening	Every instruction competes for the same attention; the ones in the middle lose.
Why the fix works	One task per turn gives each instruction full attention, and you check each result before the next.
Applies when	More than two verbs in one prompt.

5 - No example when format matters

“Match our style” describes nothing. A sample shows tone, length and structure in one move.

Symptom: facts right, voice wrong, and it stays wrong after three rounds of adjectives.

Before

Write product descriptions in our brand style.

After

Here are two of our product descriptions: [A] [B]. Write one for the item below in the same voice, length and structure.

Fix: two examples beat a paragraph of rules.

What is happening	“Punchy” means something different to every writer; a sample removes the interpretation step.
Why the fix works	Models match patterns (length, rhythm, structure) far better than they follow descriptions.
Applies when	Brand voice, fixed templates, structured output.

6 - Only saying what not to do

“Don't be boring” gives the model nowhere to go. A negative removes one option and leaves a thousand.

Symptom: the reply still does the banned thing, or overcorrects into something worse.

Before

Don't be boring. Don't use jargon. Don't make it too long.

After

Short sentences. One concrete example per paragraph. Words a 14-year-old knows. 150 words.

Fix: replace every “don't” with the “do” you actually want.

What is happening	Naming the thing to avoid puts it in front of the model, which makes it more likely, not less.
Why the fix works	A positive instruction names the target; the model aims at it instead of steering around a gap.
Applies when	Any prompt with more “don'ts” than “dos”.

7 - Assuming it knows your stuff

The model has no access to your files, your last meeting or yesterday's chat unless they are in the prompt.

Symptom: confident details about a report it has never seen.

Before

Update the Q3 numbers in the report.

After

Below is the report and a table of Q3 figures. Replace every Q2 number in section 3 with the matching Q3 figure. Report: [...] Figures: [...]

Fix: if it is not in the prompt, it does not exist. Put the key material at the start or the end.

What is happening	The model has no access to your files or earlier chats; it fills the gap with plausible fiction.
Why the fix works	Pasted material replaces guesses with facts. Placement matters because recall of the middle of a long prompt is weakest.
Applies when	Any prompt that says “the report”, “our” or “last time”.

Placement matters: models recall the beginning and end of long prompts better than the middle

After the reply Mistakes 8 to 10: what you do next

The second prompt is where most of the value is.

8 - Restarting instead of iterating

A new chat throws away everything the model has already understood. A follow-up keeps it.

Symptom: five near-identical first drafts and no second draft.

Before

[opens new chat] Write a welcome email for a budgeting app...

After

Same email, but cut the intro paragraph and make the call to action the first line.

Fix: change one thing, stay in the thread, ask again.

What is happening	The conversation is the model's working memory; a new chat resets it to zero.
Why the fix works	A follow-up keeps everything already understood and changes one variable, so you can see what that change did.
Applies when	The reply is 70% right.

OpenAI's prompting best practices put iteration first for the same reason.

9 - Taking facts on trust

Fluent is not the same as true. Statistics, quotes and citations are where models fail most confidently.

Symptom: a precise figure with no source, or a source that does not exist.

Before

Give me statistics on remote work productivity.

After

List statistics on remote work productivity with the source name and year for each. Flag any figure you are not certain of.

Fix: ask for sources and uncertainty, then check the two numbers you will actually use.

What is happening	The model predicts plausible text, not verified facts. A fabricated citation looks identical to a real one.
Why the fix works	Asking for source and confidence makes it expose what it is unsure of; a two-minute check catches the rest.
Applies when	Numbers, quotes, citations, dates.

10 - Over-prompting

Adjectives are not instructions. “World-class genius” changes the tone, not the competence.

Symptom: a 300-word prompt and a reply no better than the 30-word version.

Before

You are the world's greatest marketing genius with 50 years of experience. Think deeply and be brilliant. Write

After

You are a B2B email copywriter. Write a 4-line subject-line test for the campaign below.

Fix: a role sets vocabulary. Keep it to one line and spend the words on the task.

What is happening	“Genius” and “brilliant” shift the tone, not the capability. The model has one competence level.
Why the fix works	A one-line role sets vocabulary; the words saved go to task, context and format, which do change the output.
Applies when	More adjectives than nouns in your prompt.

Three prompts, rewritten

Real requests with the mistakes numbered, then the fixed version.

Most prompts carry two or three of these at once.

Cover letter

Scenario: applying for a product marketing role, CV in hand.

Before (mistakes 1, 2, 7)

Write me a cover letter.

After

Write a 200-word cover letter for the product marketing manager role at Acme (job ad below). Lead with my launch of a fintech app to 40,000 users. Confident, no clichés. My CV: [...] Job ad: [...]

What changed: a scoped task, the reader and role, and the two documents pasted in.

Meeting notes to actions

Scenario: 40 minutes of transcript, five people, one owner per action.

Before (mistakes 3, 4, 9)

Summarize this meeting, list the actions, write the follow-up email and tell me who was most engaged.

After

From the transcript below, list every action item as: owner, task, due date. Quote the line each one comes from. Nothing else. Transcript: [...]

What changed: one task, a fixed shape, and evidence for every item so nothing is invented.

Product description

Scenario: a new item for an online store with an established voice.

Before (mistakes 5, 6, 10)

You are a legendary copywriter. Write an amazing description for our new candle. Don't be cheesy, don't be generic, don't oversell.

After

Write a product description for the candle below in the same voice, length and structure as these two: [A] [B]. Plain words, one sensory detail per sentence, 70 words. Candle: [...]

What changed: two examples replaced the adjectives, and every “don't” became a “do”.

The 30-second checklist

Run it before you press enter. One line per mistake.

Ten boxes, thirty seconds.

1. Task

A verb and an object, with a scope.

2. Context

Who reads it, why, what happens next.

3. Shape

Count, length and structure named.

4. One job

One task per turn, or numbered steps.

5. Example

A sample pasted if style matters.

6. Positive

Every “don't” turned into a “do”.

7. Material

Files and facts pasted, key parts first or last.

8. Thread

Iterate on the reply, one change at a time.

9. Sources

Sources and uncertainty requested for facts.

10. Lean

Role in one line, words spent on the task.

Final verdict

Most bad replies are bad prompts. The fix is usually one line.

The model completes what it is given. Ambiguity in, average out. Nine of the ten fixes above add a single line: a verb, a reader, a count, an example, a source request.

• Specific beats clever. A scoped task tells the model what to do; a flattering role only tells it how to sound. Mistakes 1, 2, 3 and 10 are all the same error: words spent on style instead of scope.

• Show, don't describe. One pasted example carries tone, length and structure at once. Mistakes 5 and 6 disappear the moment there is a sample to match instead of adjectives to interpret.

• Read, change one thing, ask again. The first reply is a diagnostic, not a result. Mistakes 4, 7, 8 and 9 are all caught at this step: split the task, paste what is missing, adjust one variable, check the numbers.

• Run the checklist for a week. After that, the ten checks become reflex and the prompt is right the first time.