Why is AI giving me bad answers?

Usually the prompt is missing something: your context, a specific task or the format you want. The AI fills every gap with a generic guess. Add the missing piece and the answer often improves at once.

Should I start a new chat or keep going?

Keep going while the chat is short and on one topic. If it has become very long, switched topics or started forgetting earlier facts, start a fresh chat and open with a short summary of what matters.

How long should a prompt be?

As long as the task needs and no longer. A quick fact needs one line. A report outline or a cover letter may need a short paragraph of context plus a format. Clear and complete beats long.

Does writing in capital letters or adding “very important” help?

Rarely. Modern models follow clear, calm instructions well, and heavy emphasis can make them over-apply a rule. A specific instruction, like a word limit, works better than shouting.

Module 3 · Talking to AI: prompting · Lesson 17 of 32

Prompt mistakes and how to fix them

9 minFree GenAI course · BeginnerLesson 3.5By Saket Bhatnagar

Researched and drafted with AI, reviewed and edited by Saket Bhatnagar.

In short

Most bad AI answers come from a handful of fixable prompt mistakes, not a “dumb” model.
The big ones: a vague ask, missing context, too many tasks at once and no format.
Diagnose the symptom, fix one thing, try again. Prompting is a quick feedback loop.

Soon

Video lesson · coming soon

Prompt mistakes and how to fix them

Lala, your AI mentor, will walk you through this lesson in a short animated video. Until then, everything is here to read and try.

01 “Write my project report”
02 Think like a family doctor
03 Symptom, cause, fix one thing
04 The big four mistakes
05 Let’s play prompt doctor
06 Say what to do, not what not to do
07 Play the game, then meet your data

It is 1 am in a hostel in Lucknow. Zoya’s final-year project report is due on Monday. She types “write my project report” into an AI chat app and gets three pages about “the transformative power of technology in modern society”. Nothing about her project, a smart irrigation system.

She messages her friend Ishaan: “This AI is useless.” Ishaan reads her prompt and laughs. “It’s not useless. It has no idea what you built.” Fifteen minutes later, with a better prompt, she has a solid outline that is actually about her project.

Bad answers have causes you can find

In this module you learnt what a good brief looks like: role, context, task, example, format. You saw examples steer style, steps improve reasoning and schemas fix the shape. Almost every bad answer is one of those pieces missing or broken.

Think of it like this

Dr Bhalla’s clinic

Dr Bhalla has been the family doctor on Zoya’s street for twenty years. When a patient says “I feel bad”, she doesn’t prescribe anything. She asks: where does it hurt, since when, any fever? She names the cause, gives one medicine, and says, “Come back in three days if it hasn’t helped.” She never changes five medicines at once, because then she would not know which one worked.

In real lifeIn AI

The patient’s symptom, like a cough: In AI: What is wrong with the answer: too long, too generic, made up
Asking questions to find the cause: In AI: Checking which part of the prompt is missing or unclear
One medicine at a time: In AI: Changing one thing in the prompt, then trying again
The follow-up visit: In AI: Comparing the new answer with the old one

The prompt doctor’s loop

Drag to explore

Look at what is wrong, find the missing piece, change one thing and compare. Repeat until the answer is good enough.

The eight most common prompt mistakes

Here are the mistakes that cause most bad answers, each with the symptom you will notice and the fix. Keep this table nearby the next time an answer disappoints you.

Prompt mistakes, their symptoms and their fixes
Mistake	Symptom you notice	Fix
Vague ask	Generic, could-be-anyone answer	Say exactly what you want, for whom and for what use
Missing context	Wrong assumptions about you or your situation	Add the facts only you know: your level, your project, your limits
Too many tasks at once	Some parts skipped, the rest shallow	Split the job into steps, one task per message
No format	A wall of text, or the wrong length	State length and layout: “5 bullets”, “a table”, “under 100 words”
Only saying what not to do	It does the thing anyway	Say what to do instead: “under 80 words” beats “don’t be long”
Clashing instructions	An odd answer that half-follows both	Remove the clash, or say which rule wins
Asking for facts it can’t know	Confident, made-up details	Paste the source and ask it to answer only from that, or say “I don’t know”
One endless chat	It forgets rules from early on	Restate key facts, or start a fresh chat with a short summary

The last two link back to module 2. A model only knows its training data and what you put in the prompt, so it fills gaps with fluent guesses. And a long chat can push your early messages out of the context window, so the model truly no longer sees them.

Prompt chaining is the cure for “too many tasks”. Instead of one giant request, Zoya asks for an outline first, checks it, and then asks for one section at a time. Each step is small enough to get right, and she can correct course early.

A fix in action: Zoya’s project report

Here is the prompt Zoya started with. It has four of the mistakes in one line: vague, no context, a huge task and no format.

text

write my project report

And here is the version Ishaan helped her write. It is the first link of a chain: an outline she can check before any section gets written.

text

Role: You are a helpful guide for final-year B.Tech project reports.

Context: My project is a smart irrigation system for small farms. A soil moisture sensor sends readings to a small board, which switches a water pump on when the soil is dry. I tested it on a 10 x 10 metre plot near Lucknow for three weeks. Water use dropped by roughly a third compared with fixed-time watering.

Task: Suggest an outline for my project report. Do not write the sections yet.

Format: Chapter titles with two or three bullet points under each. Follow the usual order: abstract, introduction, literature review, design, results, conclusion.

Before you answer, ask me up to three questions about anything that is unclear.

One honest note: the AI should help Zoya plan, explain and polish, not invent her results. Her real readings go in the report, and her college’s rules on AI use still apply.

Play prompt doctor

Try it yourselfPrompt doctor: pick the best fixEach patient is a prompt with a bad result. Read the symptom and pick the treatment that fixes the real cause.

Check yourself

Your prompt gives a poor answer. You change the role, the format, the tone and the examples all at once, and it gets better. What is the problem with this approach?

Pick an answer to see why it is right or wrong.

Check yourself

Which prompt avoids the most mistakes?

Pick an answer to see why it is right or wrong.

Remember these three

Bad answers are symptoms: vague asks, missing context, too many tasks and no format cause most of them.
Say what to do instead of what not to do, remove clashing rules, and give the source for facts it can’t know.
Fix one thing at a time and compare, and split big jobs into a chain of smaller prompts.

Questions people ask

Why is AI giving me bad answers?: Usually the prompt is missing something: your context, a specific task or the format you want. The AI fills every gap with a generic guess. Add the missing piece and the answer often improves at once.
Should I start a new chat or keep going?: Keep going while the chat is short and on one topic. If it has become very long, switched topics or started forgetting earlier facts, start a fresh chat and open with a short summary of what matters.
How long should a prompt be?: As long as the task needs and no longer. A quick fact needs one line. A report outline or a cover letter may need a short paragraph of context plus a format. Clear and complete beats long.
Does writing in capital letters or adding “very important” help?: Rarely. Modern models follow clear, calm instructions well, and heavy emphasis can make them over-apply a rule. A specific instruction, like a word limit, works better than shouting.