Alex Hughes Profile picture
Sep 2 14 tweets 3 min read Read on X
MIT researchers just found out:

99% of people are prompting wrong.

They throw random words at AI and hope for magic.

Here’s how to actually get consistent, high-quality outputs:
There are 3 main ways to prompt:

👉 Zero-shot
👉 Few-shot
👉 Chain-of-thought

Each works in different scenarios.

Get this wrong, and your outputs will always be shaky. Image
Zero-shot → Just ask the question.

Example:

“Summarize this article in 3 bullet points.”

Good for simple, clear tasks.

But it collapses when the task is fuzzy or creative. Image
Few-shot → Show the model examples first.

Bad:
“Write me a cold email for a SaaS tool.”

Better:
“Here are 2 cold emails I like. Write one for [X product] in the same style.”

LLMs learn by pattern. Show → then ask. Image
This is why examples > wording.

People obsess over magic phrases (“act as an expert”).
But what matters is showing the AI the shape of the answer you want.

Show 2–3 examples, and the model locks onto the pattern.
Chain-of-thought (CoT) → Tell the AI to “think step by step.”

Bad:
“What’s the best way to grow this SaaS?”

Better:

“Think step by step:

1. Identify growth channels.
2. Evaluate pros/cons.
3. Suggest the most cost-effective one.”

Now the output has structure. Image
Here’s a practical analogy:

Prompts are like code.

If it breaks, you don’t blame the computer you debug.

→ Add an example.
→ Break it into smaller steps.
→ Reframe the ask.

Iterate until stable.
Example Prompt Debugging:

❌ Bad: “Explain quantum computing simply.”
🤷 Result: Jargon-heavy mess.

✅ Better:

“Explain quantum computing to a 10-year-old using analogies. Give 3 examples.”
This is the mindset shift:
Stop prompting like a spellcaster.
Start prompting like an engineer.

Good prompts aren’t magic words, they’re systems.
And the future? Self-improving prompts.

We’ll soon have AI that auto-tests multiple prompt variations, measures the outputs, and refines the best one.

Prompt engineering → prompt evolution.
So here’s the practical framework:

- Zero-shot for simple asks
- Few-shot for style & format
- CoT for reasoning tasks
- Debug like code when it breaks
Prompting isn’t broken.
The way most people do it is broken.

Once you apply this framework, your AI outputs will stop being random and start being reliable.
90% of customers expect instant replies.

Most businesses? Still responding days later.

Meet Droxy AI: your 24/7 AI employee:

• Handles calls, chats, comments
• Speaks 95+ languages
• Feels human
• Costs $20/mo

Start automating:

try.droxy.ai/now
I hope you've found this thread helpful.

Follow me @alxnderhughes for more.

Like/Repost the quote below if you can:

• • •

Missing some Tweet in this thread? You can try to force a refresh
 

Keep Current with Alex Hughes

Alex Hughes Profile picture

Stay in touch and get notified when new unrolls are available from this author!

Read all threads

This Thread may be Removed Anytime!

PDF

Twitter may remove this content at anytime! Save it as PDF for later use!

Try unrolling a thread yourself!

how to unroll video
  1. Follow @ThreadReaderApp to mention us!

  2. From a Twitter thread mention us with a keyword "unroll"
@threadreaderapp unroll

Practice here first or read more on our help page!

More from @alxnderhughes

Sep 1
Context length is the most important AI concept nobody explains.

It’s literally why your chatbot “forgets.”

Here’s the concept explained in plain English 👇 Image
Every Large Language Model (LLM) has a token limit.

A token = a chunk of text (≈ 3–4 characters of English).

Think of it as the AI’s working memory.

If you exceed it, the model starts dropping information.

Example:

- GPT-4o has ~128k tokens (~300 pages of text).
- Claude 3.5 Sonnet has 200k tokens (~500 pages).
- Gemini 1.5 Pro: 1M+ tokens (~3,000 pages).

But no model has “infinite memory.”Image
Why it matters:

Context length defines how much history you can pass in:

- A long chat log
- A book or research paper
- Multiple files or codebases

If your prompt + conversation > token limit → earlier parts get truncated.
Read 10 tweets
Aug 31
Holy sh*t… Claude with XML is a different beast.

Anthropic researchers quietly dropped the framework — and no one’s talking about it.

It’s like switching from calculator to supercomputer.

Here’s the structure they didn’t put in the docs:
Why XML?

Claude was trained on structured, XML-heavy data like documentation, code, and datasets.

So when you use XML tags in your prompts, you’re literally speaking its native language.

The result? Sharper, cleaner, and more controllable outputs.

(Anthropic says that XML tag prompts gets best results)Image
Why it works:

✅ Clarifies intent
✅ Mimics Claude's training structure
✅ Boosts reasoning + structure
✅ Reduces hallucination

Now let’s dive into 5 🔥 real-world use cases:
Read 11 tweets
Aug 30
You don’t need courses anymore.

Google Gemini now has 'Guided Learning' a full AI-powered tutor that explains, tests, and checks your understanding.

Here’s how it works (and why it's a game changer):
1. How to get in:

• Go to
• Start a new chat
• Choose Guided Learning
• Ask a question or upload a PDF/notes
• Turn them into a lesson with practice. gemini.google.com
2. How it works:

• Detailed explanations you can follow
• Quick quizzes and flashcards
• Pictures and YouTube videos
• Turn your documents into study sets
• Extra help when you’re having trouble
Read 7 tweets
Aug 27
The top 1% of AI users get 10x better results with the same models as everyone else.

Their secret? They mastered the skill everyone ignores: prompting.

MIT proved it drives 50% of performance.

The skill that changes everything ↓ Image
When people upgrade to more powerful AI, they expect better results.

And yes, newer models do perform better.

But this study found a twist:

Only half the quality jump came from the model.

The rest came from how users adapted their prompts.
The researchers tested this with OpenAI’s image generators: DALL·E 2 vs DALL·E 3.

~1,900 people had to recreate target images using prompts.

Result:

DALL·E 3 beat DALL·E 2 but the biggest differentiator wasn’t just the model.

It was how users changed their prompting behavior. Image
Read 13 tweets
Aug 26
I can't accept that this is real.

AI agents are doing financial analysis that would take my team days.

Through simple English conversations.

We're witnessing the death of traditional analytics.

Here's everything you need to know: Image
90% software engineers think "traditional ML" is outdated.

But it still dominates where precision matters:

• sales forecasts
• customer churn
• segmentation
• resource demand

The trick is combining that with LLMs, not replacing one with the other.
Using this setup:

→ LLM agents can call time-series models
→ Make business predictions (like demand/sales)
→ And auto-trigger actions based on output

It’s not just generative AI.

It’s predictive AI, embedded into agent workflows.

This is the core idea:

You train a forecasting model (e.g. XGBoost on SageMaker AI)
Wrap it with a simple function
Expose it to an agent as a tool
…and let the LLM decide when to use it

All via one SDK: Strands Agents
Read 10 tweets
Aug 25
Asking AI to “just do it” is not a prompt.

It’s a cry for help.

Here are 4 actual frameworks that get shockingly good results:
Today, most people prompt like this:

“Write me a marketing plan for my product.”

And then they wonder why the result feels vague, boring, and unusable.

The problem isn’t AI.

It’s your approach.
AI is like an engine.

Your prompt is the steering wheel.

But without a framework, you’re just spinning it randomly.

Frameworks turn prompting into a repeatable system.

Let’s break down 4 of the best: 👇
Read 14 tweets

Did Thread Reader help you today?

Support us! We are indie developers!


This site is made by just two indie developers on a laptop doing marketing, support and development! Read more about the story.

Become a Premium Member ($3/month or $30/year) and get exclusive features!

Become Premium

Don't want to be a Premium member but still want to support us?

Make a small donation by buying us coffee ($5) or help with server cost ($10)

Donate via Paypal

Or Donate anonymously using crypto!

Ethereum

0xfe58350B80634f60Fa6Dc149a72b4DFbc17D341E copy

Bitcoin

3ATGMxNzCUFzxpMCHL5sWSt4DVtS8UqXpi copy

Thank you for your support!

Follow Us!

:(