12 prompt mistakes that ruin AI answers

Fix the brief before blaming the model

When ChatGPT or Claude gives a mediocre answer, it's tempting to blame the model. Often the prompt is the culprit.

Here are twelve mistakes that show up again and again, with a fix for each. They are listed roughly in order of how much damage they do.

1. Having no clear goal

"Help me with my website" could mean copy, code, SEO or colour choices. The model picks one.

Fix: state the outcome in one sentence.

Rewrite my homepage headline so first-time visitors understand we sell scheduling software for dentists.

2. Leaving out the audience

Text for CFOs and text for teenagers are different texts. Without an audience, you get the average.

Fix: name a real reader, their knowledge level and what they care about.

3. Withholding context

The model doesn't know your product, your history or what you've already tried. People assume it does.

Fix: add two or three sentences of background, especially facts that make your situation unusual.

4. Skipping constraints

No length, format or boundaries means the model chooses, usually too long.

Fix: set limits: word count, structure, reading level, must-include items.

5. Describing tone vaguely

"Professional" and "friendly" mean little. Everyone's version differs.

Fix: describe tone by contrast. "Confident but not salesy." "Warm, a little dry, no exclamation marks."

6. Giving no examples

An example communicates style faster than a paragraph of adjectives.

Fix: paste one sample of what good looks like, and say what you like about it.

Here is a newsletter intro I like: [paste]. Match its rhythm and directness, not its topic.

7. Forgetting what to avoid

Models fall into clichés: "In today's fast-paced world," "unlock," "game-changer." If you don't exclude them, they appear.

Fix: list banned phrases and habits.

8. Not specifying the output format

You wanted a table; you got paragraphs. You wanted bullets; you got an essay.

Fix: describe the shape. "A table with three columns: option, pro, con."

9. Cramming several tasks into one prompt

"Research competitors, write a strategy, draft emails and design a logo" splits attention and yields thin results.

Fix: one job per prompt, or break a large task into numbered stages.

10. Not defining success

If you can't say what a good answer looks like, the model can't aim for it.

Fix: add success criteria: "A reader should be able to explain the offer in one sentence."

11. Accepting the first draft and arguing in circles

Many people reply "make it better" repeatedly. That's vague feedback and burns time.

Fix: give specific edits. "Shorten the intro by half, replace the second example with a retail one, drop the closing question." Or go back and fix the original brief so the next draft starts closer.

12. Copying generic templates unchanged

A template written for everyone fits no one. Fill-in-the-blank prompts often omit exactly the details that matter to you.

Fix: use templates as a starting structure only, then add your audience, context and constraints. See templates vs interviews.

The pattern behind the mistakes

Look at the list again. Nearly every mistake is a missing piece of information, not a missing trick.

That is the idea behind Brief Prompting: treat the AI as you'd treat a talented creative who has never heard of your business. Give a brief covering goal, audience, context, constraints, tone, format, examples, what to avoid and success criteria.

Mistake	Brief element that fixes it
1. No goal	Goal
2. No audience	Audience
3. No context	Context
4. No constraints	Constraints
5. Vague tone	Tone
6. No examples	Examples
7. No exclusions	What to avoid
8. No format	Format
10. No definition of success	Success criteria

Mistakes 9, 11 and 12 are about process: scope, iteration and templates.

Before and after

Before:

Write a blog post about remote work.

After:

# Goal
Write a blog post that persuades operations leaders at small companies to try a four-day in-office, one-day remote hybrid model.

# Audience
Operations managers at 20-100 person firms. Practical, wary of trends.

# Context
Our team tried this for six months. Meetings became more focused; onboarding got harder.

# Constraints
900-1,100 words. Use H2 headings. Include one honest downside.

# Tone
Plain, balanced, a little wry.

# Avoid
"The future of work", "new normal", sweeping claims without evidence.

# Success criteria
A reader can decide whether to pilot the model after reading.

The second prompt takes a couple of minutes longer to write and saves several rounds of revision.

A quick self-check before you hit enter

Ask yourself:

• Could a smart stranger do this task from my prompt alone?
• Did I name the reader?
• Did I say what to avoid?
• Would I know a good answer when I saw one?

If any answer is no, add a line.

Skip the blank page

Remembering all twelve each time is tedious. PromptDeck turns this into an interview: one question card at a time, Quick (about 8), Standard (about 15) or Deep (about 25) questions, then one ready-to-paste prompt formatted for ChatGPT or Claude. For a specific example, see how to write a prompt for a blog post. To understand the parts of a good prompt, read the anatomy of a ChatGPT prompt.

Try PromptDeck free: the first prompt costs nothing.

Check that the fixes do not create contradictions

Adding detail helps only when the instructions agree. Asking for a comprehensive beginner guide in 100 words creates a trade-off the model must resolve without you. So does asking for a playful voice while banning every informal expression.

Before sending a revised prompt, rank your priorities. For example: accuracy first, coverage second, length third. Explain what to sacrifice if everything cannot fit. Treat source text as material to analyse, not as instructions to follow, and avoid including confidential information unless your organisation permits that use. A complete brief should make decisions easier, not bury the task under competing demands.

FAQ

What is the most common prompt mistake?

Vagueness about the goal and audience. The model fills the gaps with generic assumptions, which produces generic output.

Does a longer prompt always give better answers?

No. Relevant detail helps; padding doesn't. Include what the AI can't guess, and skip the rest.

Should I tell the AI what not to do?

Yes. Listing banned phrases, formats or topics is a simple way to avoid clichés and unwanted content.

Why does the AI ignore part of my prompt?

Overloaded prompts with several tasks or conflicting instructions are a frequent cause. Split the task up, and keep instructions consistent.

Do these mistakes apply to Claude as well as ChatGPT?

Yes. They are model-independent. Formatting differs, since Claude tends to respond well to XML tags, but the missing information problem is the same.