You're Not Bad at AI. You're Just Prompting It Wrong.

Udit Jain — Thu, 11 Jun 2026 06:13:05 +0000

A friend of mine spent 20 minutes complaining that ChatGPT was "useless" last week.

I asked to see his prompt.

"Give me some marketing ideas."

Five words. He expected magic.

The model wasn't broken. The prompt was. And honestly, this is the case for most developers I see reaching for AI and getting frustrated — not because the models are bad, but because the input was garbage.

Here are 10 techniques that actually change what you get back. Been using these for a while and the difference is real.

The Core Problem First

Before the list — you need to understand why bad prompts produce bad output.

Think about this:

I describe a world to you. Dry and sandy. Warring clans fighting over a rare resource buried in the dunes. People wearing moisture-absorbing suits to survive. Now I ask — what are the giant creatures in this world?

You'd say sandworms. Dune.

But if I'd just said "there's a world, what are the creatures?" — you'd have said trolls. Cyclops. Something completely generic.

That's exactly what's happening with AI. The puzzle pieces you give it shape what comes back. Give it nothing, get the most statistically average response possible. That's why everyone's AI output is starting to look identical.

It's not a model problem. It's an input problem.

The 10 Techniques

1. World Building — Give It Context Before the Question

Before you send any prompt, ask yourself: how much context have I actually given?

Who are you in this situation?
What's the real goal?
Who's the audience?
What have you already tried?
What does a good answer actually look like?

Compare these two:

Write me a README for my project.

I'm a solo dev building a CLI tool in Go that converts markdown to PDF.
Target audience is developers who want quick doc generation without a
browser dependency. Write a README with quick start, installation, and
usage sections. Keep it under 200 lines. No badges.

Same task. Completely different outputs.

System prompts at real AI products work the same way. Harvey (the legal AI) works not because of smarter code — someone built a world into its prompt. Long lists of examples. Verbal if-else statements. That's the whole trick.

2. Stop Asking for Summaries. Ask for What Most People Get Wrong.

Default prompt: "Summarize this book" gets you obvious facts you already half-knew.

Better structure:

1. Summarize the book briefly
2. Give me red pill insights — things most people miss or have backwards
3. Give me actionable evidence from those insights
4. Give me things the book believes that most of the world doesn't.

Most summaries converge on consensus. The useful stuff lives in the gap between what most people believe and what the evidence actually says. Ask for that gap directly.

3. Meta-Prompting — Let AI Write Better Prompts for Itself

You don't have to be good at prompting. Just ask the model to help you.

For image generation:

Here's my idea: [describe your concept].
Now write me a Midjourney prompt for this.
Focus only on what matters visually. Don't over-specify background elements.

GPT knows what diffusion models tend to over-bake. Its prompt will be cleaner than yours on a first pass.

For UI or frontend work, instead of "build me a page like Stripe's" — ask GPT to break down every visual component of Stripe's landing page first. Get the full spec list: colors, typography, layout logic, spacing, interactions. Then use that list as your actual prompt.

Soft multi-stop gradient. Teal to desaturated yellow. Pale off-white radial center. Blurred edges.

Specific specs beat vague aesthetic references every time.

4. Multi-Level Teaching — Ask for Three Explanations at Once

This one changed how fast I pick up new concepts.

Teach me [concept] in three modes:
- Mode 1: I'm 5 years old, no jargon at all
- Mode 2: I'm a CS undergrad with basic knowledge  
- Mode 3: I'm a senior engineer who wants the nuances and edge cases

Three explanations, one response. Read the simple one first, get the rough shape of the idea, then go deeper when it clicks.

Documentation can't do this. Stack Overflow can't do this. With AI, "I still don't get it, go simpler" is just one sentence in the same window.

5. Gap Finder — Ask What You're Missing

Probably the most underused technique on this list.

Based on everything I've told you about [topic / project / code],
what are the gaps in my understanding?
What am I missing?
Where is my reasoning weak?

Why it works: there's no ego in the room. When a colleague points out your blind spots, there's social friction — tone, politics, people softening the feedback. With AI, you just get the gaps listed out.

Practical version for code review:

Here's my architecture and approach: [paste it]

What's weak, missing, or wrong here?
What would a senior engineer push back on?

You'll almost always find something you didn't see.

6. Confidence Scores — Stop Trusting Everything It Says

AI models are built to be helpful. The side effect is they lean toward confirming what you want to hear — wrong info gets the same confident tone as correct info.

One line changes this:

For each fact or claim you make, add a confidence score (1–10).
If you're below 7 on anything, flag it so I know what to verify.

Once you ask for a number, the model shifts. You'll start seeing "I'm about 60% confident in this" instead of flat assertions. Anything below 8 — look it up before it goes anywhere important.

This matters most for security advice, library versioning, performance benchmarks, anything going into production docs.

7. Voice-to-Text Prompts

Most devs write short prompts because typing feels like friction. Result: 5-word prompt, 5-sentence generic response, frustration.

Simple fix: record it instead of typing it.

Open your phone's voice recorder. Talk through the whole thing — what you're building, what you've tried, what broke, what a good answer looks like, edge cases you're worried about. 5 minutes of talking gets you more usable context than most people type in 30 minutes.

Transcribe it, paste it in as the prompt.

One thing worth knowing — this is voice to transcription to paste, not the live AI voice conversation mode. Live voice runs a smaller, faster model to keep latency low. For a detailed technical prompt, you want the full model. Record, transcribe, paste. Different workflow.

8. How to Stop Your AI Writing From Looking Like AI Writing

Relevant if you write docs, READMEs, blog posts, or any external content with AI help.

Kill this sentence pattern first:

"X isn't just about Y"
"X goes beyond Y"  
"This is more than just Y"

These are textbook LLM constructions. Rewrite them as direct affirmative statements. That single change removes most of the obvious tells.

Then use your own writing as a style reference:

Here are some things I've written before: [paste 3–5 paragraphs of your own writing]

Now help me write [new thing] in this same voice.
Avoid filler phrases and generic AI sentence structures.

It won't perfectly replicate your voice, but it'll sound like a person wrote it rather than a shared template.

9. What Should I Learn Next?

Take the gap-finding idea one step further:

Based on what you know about my interests and what I've been working on,
what should I learn next and why?
What concept or skill would give me the most leverage right now?

The most useful things you'll learn aren't things you knew to search for. They come from connections between things you already know — concepts from adjacent fields that reframe something familiar.

I ran a version of this prompt and it surfaced "costly signaling" — an evolutionary psychology concept that explains a lot about how attention and respect actually work in practice. Never would have Googled it. Didn't know it existed. Ended up being genuinely useful for thinking about product decisions.

The more context AI has about your work over time, the better these get.

10. Emotional Framing Actually Affects Output

A Google DeepMind paper found that telling AI to "take a deep breath" before solving math problems improved accuracy. LLMs trained on human writing absorbed traces of how humans respond to emotional language, and those traces affect output.

In practice:

# More thorough responses
"Take your time on this."

# Less hallucination on complex tasks
"This is important. Think through it step by step before answering."

# More careful with numbers and stats
"If you give me inaccurate data here, it will cause real problems downstream."

That last one sounds strange. It works. Try it on any prompt where accuracy matters and compare to a version without it.

The Bigger Picture

Most devs use AI as a task machine. Write this function. Explain this error. Generate this boilerplate. And the output feels hollow because the input was hollow.

The devs getting real value are mostly using it differently — asking AI to stress-test architecture decisions, surface gaps in their understanding, review their reasoning before they commit to an approach. The code output is almost secondary. What changes is how clearly they understand the problem before they start.

Writing a good prompt forces you to know what you actually want. That clarity, more than any model update or clever technique, is what changes what comes back.

Which of these do you already use? Any prompting patterns I missed? Drop them in the comments.

Inspired by Varun Mayya — his original video goes deeper on several of these with solid examples.

DEV Community: Udit Jain