Why Your AI Answers Keep Being Rubbish (And How to Fix It)

There's a moment every new AI user experiences: you type a question, get a bland wall of generic text back, and think "this AI is useless." The AI isn't useless. The prompt is. Here's what's going wrong and how to fix it in minutes.

The sandwich analogy

Imagine walking into a café and ordering "food." The server brings you... something. Maybe a sandwich, maybe a salad, maybe soup. Technically correct. But not what you wanted.

Now imagine ordering: "A toasted sourdough sandwich with melted cheddar, no tomatoes, served with a side of crisps please."

Same café. Same staff. Completely different result — because you communicated specifically.

AI works exactly this way. "Explain photosynthesis" = ordering "food." "You are a science teacher for a 9-year-old who loves Minecraft. Explain photosynthesis using a Minecraft analogy in 4 short bullet points." = the detailed sandwich order.

The 5 things missing from most bad prompts

1. A role for the AI

Starting your prompt with "You are a [role]..." changes everything. Compare:

• ❌ "Write a story about a dragon." → generic dragon story
• ✅ "You are a children's book author known for funny stories. Write a story about a clumsy dragon who is afraid of fire." → immediately more interesting

The role sets the tone, expertise level, and style before a single word of the actual request.

2. An audience

Who is this for? AI calibrates complexity and language when you specify: "explain this to a 10-year-old" vs "explain this to a university student" produces completely different answers to the same question. Always say who you're writing for.

3. Specific details

Vague = generic. Specific = useful. Instead of "give me ideas for a school project," say "give me 5 ideas for a 10-minute science presentation for 8-year-olds that doesn't require any equipment."

4. A format instruction

"As a numbered list," "in bullet points," "as a table," "in 3 short paragraphs," "as a poem" — format instructions give AI a specific structure to fill, which dramatically reduces waffle and padding.

5. Constraints

"Keep it under 100 words," "use only simple vocabulary," "don't mention [X]," "include at least one example" — constraints prevent AI from going in directions you don't want.

The 3-step prompt upgrade

Take any bad prompt and upgrade it in 3 steps:

1. Add a role: "You are a [expert type] who [key characteristic]..."
2. Add the audience: "...for [who they are, age group, knowledge level]"
3. Add format + constraints: "Write [format] of [length/count]. Include [must-have]. Avoid [don't want]."

Before: "Explain the water cycle"
After: "You are an enthusiastic geography teacher for 9-year-olds. Explain the water cycle in 4 numbered steps. Use one real-world example kids would recognise. Maximum 150 words total."

Try both versions in ChatGPT or Claude. The difference in quality will be immediately obvious.

The prompt fail activity

This is Lesson 2.1 in AI Adventures — and it's one of the most fun activities in the course. The task: deliberately write the worst possible prompt you can think of and share the hilarious result. Then rewrite it using the 5 improvements and compare. The contrast makes the lesson stick instantly — no lecture required.