12 prompt mistakes that ruin AI answers, and how to fix each one

Most disappointing AI answers come from a handful of prompt mistakes: missing context, a vague goal, no format, too many tasks at once, and trusting the output without checking. Each one has a simple fix. Below are the 12 most common, each with a bad prompt, a fixed prompt you can adapt, and one line on why the fix works.

Mistakes in what you tell the model

1. Giving no context

Bad:

Write a product description for our candle.

Fixed:

Write a product description for "Harbor Fog," a soy candle from Wick & Tide, a two-person shop in Maine. Scent: sea salt, cedar, a little smoke. Burns about 45 hours. Sold on our own site for $28. Customers are mostly gift buyers aged 30 to 50.

Why it works: the model cannot see your business, so every detail you give replaces a generic guess with a real fact.

2. Having a vague goal

Bad:

Help me with my resume.

Fixed:

Review my resume for a junior data analyst role. Goal: get past an initial recruiter screen. Point out the 5 changes that would most improve my chances, in order of impact. Resume below.
[paste resume]

Why it works: "help" has a hundred meanings, while a concrete outcome tells the model what a good answer looks like.

3. Forgetting the audience

Bad:

Explain how compound interest works.

Fixed:

Explain how compound interest works to a 15-year-old who just opened their first savings account. Use one example with small, round numbers. No formulas.

Why it works: the same topic needs completely different vocabulary, depth and examples for different readers, and the model defaults to an average reader nobody actually is.

4. Not specifying a format

Bad:

Give me ideas for a team offsite.

Fixed:

Give me 6 ideas for a one-day offsite for a 12-person remote design team. Format: a numbered list, each with a title, one sentence on why it works for remote teams, and a rough cost per person as an estimate.

Why it works: when you define the shape of the answer, you get something you can use directly instead of a wall of text to reformat.

Mistakes in how you structure the request

5. Asking for too many tasks at once

Bad:

Analyze our survey results, write a summary for leadership, draft an email to customers, and suggest three product changes.

Fixed:

Step 1 only: here are 180 open-ended survey responses. Group them into the 5 most common themes. For each theme, give the number of responses and 2 representative quotes. I will ask for the summary next.
[paste responses]

Why it works: each task gets the model's full attention, and you can correct step 1 before mistakes spread into steps 2 to 4.

6. Writing contradicting instructions

Bad:

Write a detailed, comprehensive guide to onboarding new hires. Keep it short.

Fixed:

Write a one-page onboarding checklist for new hires at a 20-person agency. Cover the first week only. Maximum 15 items, each one line. Detail can come later in a separate guide.

Why it works: when instructions conflict, the model has to guess which one you meant; removing the conflict removes the guess.

7. Not giving examples

Bad:

Write 5 subject lines for our newsletter in our brand voice.

Fixed:

Write 5 subject lines for our newsletter. Here are 3 past subject lines that performed well, so match their tone and length:
- "The bread we almost didn't sell"
- "Your Sunday starter, explained"
- "We burned 40 loaves for this"
This issue is about our new rye and a weekend baking class.

Why it works: examples show tone and length more precisely than any adjective, which is the whole idea behind few-shot prompting.

8. Leaving out length and constraints

Bad:

Write a LinkedIn post about our new hire.

Fixed:

Write a LinkedIn post welcoming Priya, our new head of operations. Under 120 words. No hashtags, no emoji, no "thrilled to announce." Mention her previous work in logistics and one thing she will own in her first 90 days: our supplier process.

Why it works: constraints rule out the generic defaults you would otherwise spend time deleting.

Mistakes after you hit enter

9. Accepting the first draft

Bad:

Thanks, that's fine.

Fixed:

This is close. Three changes: the opening is too slow, so start with the customer's problem. Cut the second paragraph. Make the call to action specific: book a 15-minute call, not "get in touch." Keep everything else.

Why it works: the first answer is a draft, and specific feedback on what to keep and what to change usually gets you most of the way in one more turn.

10. Not asking it to ask you questions

Bad:

Write a cover letter for this job.
[paste job posting]

Fixed:

I want to write a cover letter for the job below. Before writing anything, ask me up to 6 questions about my experience and motivation that would make the letter specific to me. Wait for my answers.
[paste job posting]

Why it works: the model knows what information a strong answer needs, so letting it interview you fills gaps you did not know were there.

11. Trusting facts without checking

Bad:

List the main regulations a food truck in Austin needs to follow.

Fixed:

List the categories of permits and regulations a food truck in Austin typically needs (health, fire, business license, parking). For each, tell me which type of official body usually handles it so I can verify on their website. Mark anything you are unsure about as [VERIFY].

Why it works: language models can state outdated or wrong facts with full confidence, so asking for categories plus where to verify turns the answer into a checklist instead of a source.

12. Pasting sensitive data

Bad:

Here is our full customer export with names, emails and order history. Find our best customers.
[paste 2,000 rows]

Fixed:

I have a customer export with these columns: customer_id, first_order_date, order_count, total_spent, last_order_date. Here are 5 example rows with fake values. Write a formula or short script that ranks customers by value and flags anyone who has not ordered in 90 days. I will run it on the real data myself.

Why it works: the model only needs the structure of your data to help, and keeping real personal or confidential information out of the chat avoids privacy and contract problems entirely.

Depending on the tool, your plan and your settings, conversations may be stored or reviewed. Check your company's AI policy before pasting customer data, contracts, credentials, or unreleased financials, and remove names and identifiers when you can.

A 30-second pre-flight check

Before you send any prompt that matters, run through these questions:

1. Does the model know who I am, what this is for, and who will read it?
2. Is the goal stated as a concrete outcome?
3. Did I specify format and length?
4. Is it one task, or should I split it?
5. Do any instructions contradict each other?
6. Would one example make the target clearer?
7. Should the model ask me questions first?
8. Is there anything in here I should not be pasting?

If you want to go deeper into why these fixes work, prompt engineering for beginners walks through the core techniques, and how to write better ChatGPT prompts gives a reusable template.

A reusable template that avoids most of these mistakes

Mistakes 1 to 8 disappear when you fill in the same five parts every time. Save this and adapt it:

ROLE: You are [type of expert relevant to the task].
TASK: [one clear task, stated as an outcome].
CONTEXT: [who I am, what this is for, key facts, who the audience is].
FORMAT: [structure, length, tone].
CONSTRAINTS: [what to avoid, what must be included, what to mark as uncertain].
Before you start, ask me any questions you need answered to do this well.

Different assistants respond a little differently to the same structure; ChatGPT vs Claude vs Gemini prompting covers the differences worth knowing.

If filling in the template by hand gets repetitive, a prompt generator such as Promptay builds this structure from one sentence about what you need, and you edit the result before sending it.

Frequently asked questions

What is the most common prompt mistake?

The most common mistake is giving no context. People type a short request as if the model already knows their business, audience and goal, and get a generic answer back. Adding three or four specific facts about the situation usually improves the answer more than any other single change.

Why does ChatGPT give generic answers?

Generic answers usually come from generic prompts. When the prompt lacks context, audience, format and constraints, the model fills the gaps with the most typical response it can produce. Give it specific details, an example of what you want, and clear limits, and the output becomes specific too.

Are longer prompts always better?

No. Longer prompts help when the extra words add context, examples or constraints the model needs. Padding, repeated instructions or contradicting requirements make results worse. Aim for a prompt that contains everything needed for the task and nothing else.

How do I stop AI from making things up?

You cannot fully prevent it, but you can reduce it and catch it. Provide the source material yourself, ask the model to mark anything it is unsure about, and tell it to say when it does not know. Always verify names, numbers, dates, laws and citations against an official source before using them.

Is it safe to paste work documents into an AI chat?

It depends on the tool, your plan, your settings and your employer's policy. Many companies restrict what can be shared with AI assistants. When in doubt, remove names and identifiers, share only the structure or a sample with fake values, and check your company's AI policy first.