Your AI keeps giving you mush. The prompt isn't bad. It's broken in three specific places.
If you've ever typed a question into an AI chat, read the answer, and thought "this is technically correct and yet completely useless," you're not alone. The instinct is to blame the model ā newer, smarter, more expensive. But almost every time, the model isn't the bottleneck.
The prompt is.
And the prompt is broken in the same three ways for almost everyone.
The new-hire test
Picture training a new hire. If you said "do good work" and walked away, you'd get nothing useful back. Not because they're bad at their job. Because you gave them no audience, no goal, no examples, no shape.
Most prompts read exactly like "do good work."
Generic instructions. No examples. No clear shape. And then we're surprised when the output is generic.
The fix is unglamorous: three small changes to how you write the request. None of them require a new model, a paid tier, or a course on "prompt engineering." Five minutes. Different output forever.
š„ Prefer to watch this?
If you'd rather have this delivered as a three-minute video ā same lesson, calmer pace, on the Learn in 3 channel ā here it is:
šŗ Watch on YouTube: 3 Prompt Mistakes That Wreck Your AI Prompts
Otherwise keep reading. The text version below is the whole thing, end to end.
Mistake 1 ā The vague ask
The most common mistake, by a wide margin.
"Write me a blog post."
That isn't a prompt. It's a wish. The AI doesn't know your audience, your length, your tone, or your goal ā so it averages everything and hands you the middle. Bland marketing copy. Cold and forgettable.
The fix is unglamorous. Add the audience. Add the length. Add the tone.
Before:
Write me a blog post.
After:
Write a 600-word blog post for marketing managers,
in a friendly editorial tone, on prompt engineering basics.
Specifics aren't extra. Specifics are the prompt.
The difference is bigger than you'd expect ā go try the two versions side by side, on whichever AI chat you use.
Mistake 2 ā No role assigned
Asking the AI a question without telling it who to be is asking a stranger on the street. You'll get an answer, but it'll be the median answer.
Tell it to answer as a marketing director. Or a kindergarten teacher. Or a friendly editor. Or ā if you're a developer ā a senior Python engineer reviewing a junior's PR.
Same question, different expert, completely different answer.
Before:
How do I make this function faster?
After:
You're a senior Python engineer doing a code review.
The function below is O(nĀ²). Rewrite it for O(n) where possible,
and explain the change in two sentences.
The role unlocks the model's reservoir of "what would this kind of expert say." That reservoir is huge. It just sits dormant until you turn it on.
Pick the role first. Then ask.
Mistake 3 ā No examples
You're asking the AI to match your style without ever showing it your style. That's guessing.
Paste two or three short samples of the kind of output you want. The AI copies the pattern. This trick is called few-shot prompting, and three examples beats a paragraph of instructions every time.
Three is the sweet spot. One looks like a coincidence. Two leaves wiggle room. Three creates a clear pattern the model locks onto.
Before:
Write a product description in our style.
After:
Write a product description in our voice.
Here are three of ours we like:
1. "Cedar-aged. Quietly defiant."
2. "Hand-bound. Mistakes welcome."
3. "Loud in the right places."
Now write one for a leather-bound notebook.
You'll watch the style transfer happen in real time. Sentence length, vocabulary, rhythm ā all of it.
The trap most people fall into is picking three examples that contradict each other. One formal, one casual, one short, one rambling. The AI gets confused and gives you mush.
Choose three examples that share the same voice and structure. Consistency is the lesson.
The fix, all in one message
Here's what it looks like when you stack all three together:
You are a friendly editor.
Rewrite this paragraph for a tired Tuesday-morning reader,
in under 80 words.
Here are three short rewrites I've loved:
1. [example one]
2. [example two]
3. [example three]
Here's the paragraph: [...]
One message. A role. A goal. A constraint. Three examples.
The AI has almost nothing left to guess. The output sharpens immediately ā usually on the first try.
That's it. That's the whole framework.
Try it in the next five minutes
Don't bookmark this. The lesson dies in the saved tabs.
- Open your last AI chat.
- Pick a prompt that disappointed you.
- Rewrite it with a role, a clear goal, and three examples.
- Send it again.
That's the experiment. Five minutes, max.
If the output sharpens ā and it will ā leave the before-and-after in the comments. I read all of them and reply with the next thing to tighten.
Originally published as part of Learn in 3 ā one idea, broken into three, every week. ā https://youtube.com/@learn-in-3
Top comments (0)