Search for a command to run...
No comments yet. Be the first to comment.
Privacy isn't dead. It's just being rebuilt. The old internet promised us "move fast and break things." We moved fast. We broke things. Now we're realizing what we broke: privacy. Every LLM you use wa
Coding is changing faster than you think. Not just "AI helps with autocomplete." AI is becoming the primary developer. You become the reviewer. The Skill Inversion Old model (2020): Human writes code.
Single AI agents are limited. They make mistakes. They miss context. They take wrong paths. Multi-agent systems don't have these problems. They have different problems. But better ones. The Single Age
You've learned basic RAG. You embed documents. You retrieve them. You feed them to an LLM. It generates an answer. Then you deploy it. And it breaks. Because the real world isn't simple retrieval. Adv
The biggest AI models aren't winning anymore. GPT-4: 1.8 trillion parameters Claude 3 Opus: 200+ billion parameters Claude 3.5 Sonnet: Unknown but smaller Llama 3.1: 405 billion parameters PhiMoe (ope
Muhammad Zulqarnain | Full Stack AI Engineer & Geospatial Developer
16 posts
A blog by Muhammad Zulqarnain — Full Stack AI Engineer & Geospatial Developer based in Turku, Finland. I write about RAG systems, LLMs, Prompt Engineering, Next.js, TypeScript, and geospatial development. Practical insights, deep dives, and real-world AI solutions.
The era of "magic prompts" is over.
You know the ones. The thousand-word prompts with perfect formatting. The ones everyone shared on Twitter like they were gold.
"Use this system prompt and get 10x better outputs!"
Then models got better. And those prompts stopped working.
Welcome to Prompt Engineering 2.0. It's completely different.
2023 (Prompt Engineering 1.0)
Prompts were the bottleneck.
Model quality was mediocre. Model capability varied wildly. You needed to engineer every prompt carefully:
Use specific formatting
Add examples
Use specific magic words ("think step by step", "let's think")
Avoid certain phrasings
Test hundreds of variations
Small prompt changes = big output changes.
2025 (Prompt Engineering 2.0)
Models are the bottleneck.
Models are so capable now that prompts matter way less.
Try this: Prompt 1: "What's 2+2?" Claude 3.5 Sonnet: "4"
Prompt 2: "What is the answer to this mathematical equation: two plus two, expressed as a numerical value?" Claude 3.5 Sonnet: "4"
Same output. The magic prompt tricks don't work because the model is already smart enough.
1. Clarity > Cleverness
No longer: "You are an expert AI assistant with 20 years of experience..."
Now: "What is 2+2?"
Models understand simple, direct prompts better than flowery language.
The fancy system prompts? They add noise.
2. Structure > Formatting
Old way:
Instructions:
1. ...
2. ...
3. ...
Context: ...
Task: ...
Output format: ...
New way:
Fetch user data
Analyze sentiment
Return top 3 insights
Why? Modern models understand intent without the scaffolding.
3. Few-shot Learning > Prompt Engineering
Old way: Spend hours perfecting a single prompt.
New way: Show 2-3 examples and let the model generalize.
Example: "Convert these to JSON:
Input: John, 25, Engineer Output: {"name": "John", "age": 25, "role": "Engineer"}
Input: Sarah, 32, Doctor Output: {"name": "Sarah", "age": 32, "role": "Doctor"}
Now do this: Mike, 28, Designer"
One example (few-shot) is better than a page of instructions.
4. Tool Use > Complex Prompts
Old way: Try to get the model to do everything in one prompt.
New way: Give the model access to tools.
The model decides:
What tools to use
In what order
What parameters to pass
When to stop
Prompt becomes simple: "Fetch the user's order history and return the most expensive item."
Model: Calls database tool, filters, returns result.
5. Chain of Thought (Explicit) > Chain of Thought (Implicit)
Old way: "Think step by step" in the prompt.
Now: Use structured reasoning with extended thinking.
Prompt: "What's the best strategy for this chess position?" Model: Spends compute time (5-30 seconds) thinking deeply. Model: Returns answer with reasoning shown.
Computationally more expensive. Vastly better results.
1. Context Quality
Garbage in, garbage out. Still true.
But now the focus is:
Are you giving the right context?
Is the context accurate?
Is it relevant to the question?
Not "did I phrase the question perfectly."
2. Task Clarity
What are you trying to do?
If the model is unclear on the goal, no clever prompt trick will help.
But if the goal is clear? Even a mediocre prompt works.
3. Output Format
This still matters. Models are good at:
JSON
CSV
Markdown
XML
Specify the format. Don't need to spend hours perfecting it.
4. Model Selection
This is the new prompt engineering.
Choosing the right model > choosing the right prompt.
GPT-4o for vision? Claude 3.5 Sonnet for reasoning? Gemini 2.0 for multi-modal?
The model choice is the strategy.
5. Temperature & Parameters
Done right, this is more important than prompt wording.
temperature=0: Consistent, predictable temperature=1: Creative, diverse max_tokens: Control output length top_p: Control diversity
Tweak these. Not the prompt.
Old way: Write one perfect prompt. New way: Iterate with real data.
The Process
Write a simple prompt
Test on 10 examples
See where it fails
Add context for failure cases
Test again
Repeat until good enough
It's not about finding "the perfect prompt."
It's about finding "what context makes this task clear."
Prompt Engineer (Old Definition)
"I specialize in writing perfect prompts."
This job is dying. Models are too good.
The New Prompt Engineer
Understand:
What the model can do
What it can't do
How to structure tasks for reliability
How to evaluate quality
How to iterate on context
How to chain models together
How to use tools effectively
Different skill set. Higher leverage.
1. Structured Prompting
Learn how to structure complex tasks:
Input format
Processing steps
Output format
Error handling
2. Few-Shot Learning
How to give examples. How many. What kind.
3. Tool Integration
How to design prompts that leverage tools effectively.
4. Evaluation
How to measure if a prompt actually works.
Not "does it look good" but "does it solve the problem."
5. Model Comparison
Test your task on multiple models. See which wins.
Not all models are equal for all tasks.
Prompts will become:
Shorter
Simpler
More about context than creativity
More about structure than wording
The engineers winning will be those who:
Understand model capabilities deeply
Can design effective task structures
Know how to evaluate quality
Can integrate tools and workflows
The ones losing will be those who:
Sell "magic prompts"
Focus on prompt wording
Believe in prompt formulas
Don't understand the underlying models
Dying: "Prompt Engineer" as a pure role Thriving: Engineers who understand AI + can design systems
The future isn't prompt specialists. It's full-stack AI engineers who can:
Choose models
Design prompts
Integrate tools
Evaluate outputs
Iterate to reliability
That's where the $200K+ salaries will be.
Not in perfecting a single prompt.niques Are Dying