TL;DR

AI models generate text based on probability, which can lead to confidently incorrect answers. Prompt engineering solves this by providing specific, structured instructions so chatbots deliver exactly what you need.

Core Concepts:

• System vs. User Prompts: Set background rules and personas first (System), then assign the immediate task (User).

• Zero-Shot vs. Few-Shot: Don’t just tell the AI what to do. Provide 2-3 high-quality examples (Few-Shot) to guarantee the correct style and format.

The 4 Golden Rules:

1. Ordering: Place your most critical instructions at the very beginning or the very end of your prompt. AI might miss things buried in the middle.

2. Be Hyper-Specific: Define exact word counts, formatting (e.g., “3 bullets, under 20 words each”), and the target audience.

3. Chain Your Tasks: Break complex workflows (research, outline, write, edit) into separate, sequential prompts rather than one massive request.

4. Provide a “Golden Sample”: Give the AI a flawless example of your previous work to mimic for tone and subjective tasks.

Merhaba👋🏻

I am a Software Engineer with 10+ years of experience. My goal is to close the gap between the technical and the non-technical, making AI accessible to everyone, regardless of their background.

Subscribe now

This is a story about prompt engineering, and why getting it wrong nearly ruined a surfboard, and a vacation.

A friend of mine is really into surfing.

He loves going to all around the world, surfing through the oceans, and getting better and better each time. And he is also stubborn: he likes to push himself to go higher level each time.

Recently, he went to a surf camp in Morocco: more of a mid/intermediate level of surfing area. Before he went there, he asked Claude to recommend a surfboard for his skill level. Claude gave him a confident suggestion. He bought the board, packed it, flew to Morocco, and paddled into the ocean.

The board was a wrong suggestion for him. It was not really appropriate for his skill level yet, and it was hard for him to control. In the end, surf instructor came to rescue him, and he survived.

When he got back to hotel, he asked Claude why it suggested this surfboard and he was almost dying!

Claude said “I am sorry, you are right”.

Although AI models are quite smart these days, it is still up to us to know how to talk to them to get what we want.

This is why I am writing this article: to talk about how to effectively communicate with AI, aka Prompt Engineering.

Subscribe now

What Is an LLM? The AI Behind ChatGPT, Claude, and Gemini

Before we get into prompt engineering, I want to briefly talk about what LLM is. As you may heard already, chatbots like ChatGPT, Claude, Gemini are considered as LLM (Large Language Model). By definition, Language Model is a model that holds statistical information about one (or multiple) language(s). This way, it can predict how to complete a sentence based on the given context. For example;

“A bike has 2 ___” will likely be completed by Language Model (LM) with “wheels” instead of “green”. That’s why LMs considered as output generator based on probability. They can generate infinite number of new contents, that’s why we see Generative AI term everywhere.

You probably also heard of token, which is the basic unit that describe a language model. I wrote an article about what is token, and how tokenization works:

LLMs, on the other hand, mainly refer to models that trained on large datasets. It usually measured in parameters. Parameters numerically translate human language into patterns the model can understand. Here is the comparison of different GPTs:

As you see, GPT-3 is roughly 1500 larger than GPT-1. The reason we see this jump is to make model smarter to increase its capabilities. This way, you can summarize text, translate it into another language, or get help with coding.

As I described above, language models, by nature, take some text and try to complete it with a statistical approach. However, starting with ChatGPT, followed by Claude, Gemini and many more, LLMs have been trained as a chatbot. This process called fine tuning: further training a model for a specific task. That way, instead of being a completion machines, they answer questions.

Today, we will talk about how you can maximize the performance of an AI chatbot for your specific needs by using Prompt Engineering techniques.

Leave a comment

What Is Prompt Engineering? (And Why Everyone Needs It Now)

Prompt is a text that contains instruction for AI to perform a task. It could be a simple question like “What is the height of Empire State Building?” or a more complex task like building a website.

Prompt engineering is a process to tweak AI with instructions to generate a desired response.

This sounds technical, but it actually more like briefing a new hire. You wouldn’t just say “Make a good marketing plan”, yet many people do exactly this to AI.

When you are communicating with AI, describe what you want to achieve specifically. AI chatbots are pretty good at following instructions, so instead of asking “Write a marketing brief for X product”, ask it “Write a marketing brief for X product under 5 sentences and use a formal tone”.

System Prompt vs. User Prompt: The Hidden Setup Most People Skip

Before we get into the templates, there is a foundational setup that most people are missing. When you open an AI tool, there are actually two layers of communication:

1. The User Prompt: What you say in the moment (the direct task).

2. The System Prompt: The background rules you set before the conversation even starts (the job description).

Let’s say you want to write a report with the same format every time. And you want the chatbot act in a certain way with role playing.

You can add its role to system prompt:

You are an executive communications specialist who turns 
raw working notes into clear, decision-ready internal reports. Your job is to turn messy 
notes into polished internal reports. My company uses this structure: 
Executive Summary → Key Findings → Recommended Actions → Next Steps. 
Always use this structure unless I say otherwise. 
My audience is non-technical senior management. Avoid jargon. 
Use bullet points for findings, full sentences for recommendations. 
Tone: confident, not academic.

And give its task with user prompt:

My notes from last week's quarterly review. It was a mess, help me make sense of it. 
We missed our target but retention was actually up.

[paste notes]

If you are using ChatGPT, you can add your system prompt as Custom Instruction under the Settings

Zero-Shot vs. Few-Shot Prompting: Why Examples Beat Instructions Every Time

You can try to explain what you want and be very specific in your prompt, but you may still not be very happy with AI’s response. Although it is still important to be specific, there is also another method called in-context learning.

In-context learning basically means giving correct examples to AI. Each example that is provided is called shot.

If there is no example in the prompt, it is called zero-shot prompting.

“Summarize this meeting transcript in three bullet points.” (Works fine for basic, low-stakes tasks).

If there are several examples, it is called few-shot prompting. You show the AI what a good output looks like before asking it to do the work.

Instead of hoping the AI guesses your style, you give it a baseline:

Here are examples of my notes vs. the update I want:

Example 1:
Notes: "fixed that annoying login bug, took forever, finally done"
Update: "Resolved a persistent authentication issue that was 
blocking user onboarding. Fully deployed and tested."

Example 2:
Notes: "meeting with Sarah about Q3 roadmap, nothing decided yet, 
she wants more data"
Update: "Held a roadmap alignment session with Sarah (Q3). 
Next step: gather additional data to support prioritization 
decisions — targeting next week."

Example 3:
Notes: "helped Mehmet debug his API thing, not my job but whatever"
Update: "Provided cross-team support on an API integration 
issue. Resolved same day."

Now do this one:
Notes: "spent half the day in pointless meetings, did manage to 
finish the supplier comparison doc nobody asked for"

Few-shot prompting is especially useful when you want to get consistent response from AI.

Subscribe now

The 4 Golden Rules of Prompt Engineering for Better AI Responses

Through trial, error, and testing, these four principles cover about 80% of the improvements you’ll ever need:

• Ordering matters: AI models pay the most attention to the very beginning and the very end of a prompt. If you bury a crucial instruction in the middle of a long paragraph, the AI is likely to miss it. Put your most critical rules first or last.

• Be embarrassingly specific about formatting: Don’t just ask for a summary. Ask for “a 3-bullet summary, each bullet under 20 words, written for a non-technical stakeholder”. Less guessing equals fewer rewrites.

• Break complex tasks into steps: Don’t ask the AI to research, outline, write, and edit a document all in one go. Sequence it. Ask for the outline first. Review it. Then tell it to write section one.

• Provide a “Golden Sample”: If a task requires subjective judgment or a specific creative voice, give it a high-quality example to mimic.

AI is very good at following instructions but you have to tell it exactly what you want. Make sure to give it clear instructions with examples, and provide necessary context. However, it is still a bit of a trial and error: you probably need to refine your prompts until you get you want.

Which part clicked for you? Or which part still feels fuzzy? Drop it below; your question might be the next article!

Leave a comment

PS: If you're new here and wondering why a software engineer is writing about all this - here's why I started Becoming with AI."

TL;DR:

AI “coworkers” aren’t chatbots you ask questions. They’re persistent agents that proactively do work alongside you; scheduling, monitoring, drafting, flagging. The data says they boost efficiency by ~20%, not 10x. The biggest gains come from augmentation, not from replacing AI with humans.

I believe, the future isn’t humans vs. AI. It’s humans + AI teams. This article breaks down what that actually looks like, what the real concerns are, and how to prepare without panicking.

Welcome👋🏻

I am a Software Engineer with 10+ years of experience. My goal is to close the gap between the technical and the non-technical, making AI accessible to everyone, regardless of their background.

Subscribe now

It’s Monday morning. You open your laptop, coffee in your hand, bracing for the usual chaos. But your calendar is already organized. Your emails are prioritized. Your first meeting has a prep doc waiting in your inbox.

You didn’t do any of this.

Your AI coworker did. Over the weekend. While you were doing something actually human, like ignoring your phone at Easter brunch.

This isn’t science fiction anymore. And it’s not even next year. Some companies like Salesforce is already deploying agents that handle real workflows; delivering agentic IT and HR solutions. And the shift from “AI will replace us” to “AI will work with us” is happening faster than most people realize.

Here’s what that actually means for your day-to-day, and what to do about it.

What Is an AI Coworker? (And Why It's Not a Chatbot)

Let me save you some time: an AI coworker is not a chatbot you type questions into.

We call this an “AI agent”, but it is a digital teammate that proactively does work on your behalf. The difference matters. A chatbot waits for you. An AI coworker shows up with work already done.

There are three flavors worth understanding, and they sit on a spectrum of independence:

1. Task Agents: These handle specific, repeatable workflows. Scheduling meetings, entering data, pulling research. Think of them as a very reliable intern who never forgets a step.

2. Collaborative Agents: These work with you in real-time. Coding alongside you, drafting copy while you edit, analyzing data while you interpret. You’re both in the kitchen at the same time.

3. Persistent Agents: These are always on. Monitoring your campaigns, watching your inbox, flagging anomalies at 2 AM. They don’t clock out.

The thing most people miss is that these agents don’t replace a person in a chair. They replace the tasks that were eating that person alive. That’s a different thing entirely, and it’s where we need to focus.

What the Productivity Data Actually Shows: AI Augmentation vs. Replacement

Here’s what’s actually happening in the numbers:

According to recent Anthropic report, software developers lead the pack, accounting for 19% of the total gains. The rest of the top five might surprise you: operations managers (6%), marketers (5%), customer service reps (4%), and high school teachers (3%).

I know. That sounds underwhelming compared to the “10x productivity” headlines. Good. Because realistic expectations are the single greatest defense against disappointment.

~20% productivity gain is not a small number. If you spend 10 hours a week on finding a bug, writing a documentation, analyzing a different code base, a 20% efficiency gain hands you back 2 hours. Every week. That’s 100 hours a year. That’s 2,5 weeks of working full time redirected toward things that actually require your brain.

But here’s the key insight that most of the hype-cycle articles skip: the biggest gains come from augmentation, not from replacement. The companies seeing real results aren’t firing people and plugging in AI. They’re giving their existing people AI teammates and watching what happens when humans spend more time on judgment, strategy, and creativity.

The AI does the prep work. The human makes the call.

That’s the model. It’s not flashy. But it works.

AI Coworker in Practice: 3 Real-World Scenarios

Theory is nice. Let me show you three scenarios that make this concrete.

Scenario A: The Marketing Manager

Your AI coworker monitors campaign performance overnight. While you sleep, it spots that your Facebook ad spend is burning cash on an audience segment that stopped converting two days ago. It flags the anomaly, suggests reallocating budget to your top-performing Instagram set, drafts a morning summary report, and schedules a team sync for 9:30 AM.

You walk in, review the recommendation, and make the call. Yes, shift the budget. No, hold off on pausing that experiment.

The AI did the monitoring and the math. You did the judgment.

Scenario B: The Software Engineer

This one’s personal for me. Tools like Claude Code already handle boilerplate and debugging. The AI writes the repetitive scaffolding; the API endpoint that looks like the last 15 you built. It catches the null pointer exception you would have spent 40 minutes hunting.

The human focuses on architecture and user experience. The “why are we building this” and “how should this feel” questions that no model is answering well yet.

Bloomberg reported on a “productivity panic” in engineering; are we just coding faster, or are we coding better? I think about this a lot. Faster is easy to measure. Better is what matters. And “better” still requires a human who understands the problem deeply enough to know when the AI’s suggestion is technically correct but architecturally wrong.

Speed without direction is just expensive chaos.

Scenario C: The Small Business Owner

You run a local service business. Your AI coworker handles customer inquiries at 11 PM when someone finds your website after hours. It books appointments based on your real-time availability. It sends you inventory alerts when stock drops below your threshold.

You focus on the things that actually grow a small business: relationships, reputation, and showing up as a human in a full of automated competitors.

The pattern across all three scenarios is the same: AI handles the “keeping the lights on” work. You handle the “deciding where to point the lights” work.

Will AI Replace My Job? The Real Concerns, Addressed Honestly

I’m not going to pretend this is all upside. There are real tensions here, and dismissing them would be dishonest. So let’s walk through them plainly.

Job Displacement: “Will I lose my job?”

Some roles will shrink. That’s the honest answer. Data entry, basic report generation, first-tier customer support; these are already being absorbed by agents. But new roles are emerging alongside them: AI ethics consultants, automation architects, AI engineers, people who manage and audit AI systems.

The pattern isn’t new. Spreadsheets didn’t eliminate accountants. They eliminated the version of accounting that was purely manual calculation, and created space for analysis and advisory work. The transition is real, though. And pretending it’s painless would be a lie.

Skill Atrophy: “If AI does everything, do I lose competence?”

This is the “calculator debate” for the AI age. When calculators arrived, people worried we’d forget how to do math. We did, sort of. But we gained the ability to solve problems that were previously impossible because we weren’t spending all our time on long division.

The risk is real. If you never write a first draft yourself, your writing muscle weakens. If you never debug manually, your diagnostic instinct fades. The answer isn’t to avoid AI; it’s to be intentional about which skills you keep sharp and which ones you deliberately offload.

The Productivity Trap: “Am I just expected to do more?”

This one keeps me up at night. If AI saves you 2 hours a week, does your company give you breathing room, or do they fill those 2 hours with more output expectations?

Historically, productivity tools have been used to increase throughput, not decrease working hours. I don’t have a clean answer here. But I think the people who set boundaries early; who say “I’m using AI to do better work, not just more work”; will be the ones who actually benefit.

Transparency: “How do I know what it did or why?”

When your coworker is human, you can ask them to walk you through their reasoning. When your coworker is an AI, you need audit trails, logs, and a habit of checking the work. This is the management skill most people haven’t built yet. You need to learn to review AI output the way a senior engineer reviews a junior engineer’s code: trust but verify, every time.

Subscribe now

How to Prepare for AI Coworkers: 4 Practical Steps

Here’s what I’d actually do if I were starting from zero today. These are four shifts in mindset, each with a concrete step attached.

1. How to Delegate Tasks to AI (Like a Manager, Not a User)

This is the skill that separates people who get value from AI and people who don’t. Delegating to AI is exactly like managing a junior employee: you need to give clear instructions, set expectations, and check the work.

Most people either over-trust (hand everything off and never review) or under-trust (micromanage every output and waste the time savings). The sweet spot is structured delegation.

Here’s a prompt to practice this right now:

I need you to act as my Monday morning prep assistant. 
Here are my priorities for this week: [paste your top 3-5 priorities]. 
Review the following emails [paste or describe your inbox] and: 
(1) flag anything that’s urgent and related to my priorities, 
(2) draft a one-paragraph summary of what needs my attention today, 
(3) suggest one thing I can defer to Wednesday or later. Be direct. No filler.

Paste that into Claude or Gemini with your actual priorities and inbox summary. See what comes back. Adjust the instructions based on what was useful and what wasn’t. That adjustment process? That is the skill.

2. The Skills AI Can't Replace: Why Human Judgment Still Wins

Critical thinking. Creativity. Empathy. Ethical reasoning.

These are the skills that become more valuable, not less, when AI handles the routine work.

The question to keep asking yourself: “What am I doing right now that requires me to be a human”? If the answer is “nothing”, you’re doing AI’s job. Swap.

3. How to Experiment with AI Tools Before Your Company Makes You

Experiment with tools before they’re mandated by your company.

If you want to try this today:

1. Pick one repetitive task you do every week (email sorting, meeting prep, report drafting).

2. Open Claude or Gemini.

3. Describe the task in plain English and ask the AI to do it with a sample input.

4. Evaluate the output: What was good? What was wrong? What instructions would you change?

5. Run it again with better instructions.

That’s one cycle. Do five of those and you’ll understand more about AI collaboration than most “AI strategy” workshops teach in a full day.

4. The One Question That Reframes Your Relationship with AI

When your company inevitably adopts AI tools, don’t ask “Is this going to replace me”? Ask: “What will I do that the AI can’t”?

That question reframes you from a potential victim to the person defining the human-AI split on your team. It’s a power move disguised as a question.

The Human Element: Why Managing AI Is Still a Deeply Human Skill

Here’s what I keep coming back to: AI coworkers aren’t coming for your job. They’re coming to your job. There’s a difference, and it matters.

A coworker who handles your Monday morning chaos so you can walk into work thinking about strategy instead of a repetitive work? That’s not a threat. That’s a teammate. A persistent agent that catches the budget anomaly at 2 AM so you can make the call at 9 AM? That’s not replacement. That’s augmentation.

But here’s the part I’m still working through myself: the line between “AI handles the grunt work” and “AI handles so much that I forget how to do my own job” is blurry. It moves. And it requires honest, ongoing attention.

I believe, the people who will thrive in this shift aren’t the ones who resist AI and aren’t the ones who surrender everything to it. They’re the ones who learn to manage it; like a coworker, not like a magic button.

Clear instructions. Regular check-ins. Honest evaluation of the output. That’s management. And it turns out, managing an AI teammate well is a deeply human skill.

Two questions if you’re up for it:

1. What’s the one recurring task in your week that you’d hand to an AI coworker tomorrow if you could? What’s actually stopping you from trying it today?

2. Have you already hit the “productivity trap”; where AI saved you time but your workload just expanded to fill it? How did you handle it?

Leave your comment below, I read everything.

Leave a comment

PS: If you're new here and wondering why a software engineer is writing about all this - here's why I started Becoming with AI."

TL;DR:

What AI agents are: Not sentient robots, not AGI - just a simple formula: Brain (an LLM like Claude) + Hands (tools like search or a calculator) + Loop (the ability to keep trying until the job is done).

How AI agents work vs. chatbots: The difference isn't tools - modern chatbots like Claude already use them. It's who drives the loop: with a chatbot, you prompt every step. An agent takes a goal, chains its own steps, checks its own work, and finishes the task without you prompting it each time.

When you actually need one: Use Claude for quick answers. Build a custom AI agent when you want to automate a multi-step process using your own data - something that runs in the background while you sleep.

Welcome👋🏻
I am a Software Engineer with 10+ years of experience. My goal is to close the gap between the technical and the non-technical, making AI accessible to everyone, regardless of their background.

Subscribe now

Everyone’s talking about AI agents. Your LinkedIn feed is likely drowning in posts about “autonomous AI agents that will replace your entire team.” VCs are throwing money at anything with the word “agent” in the pitch deck. Every second tutorial starts with: “In this revolutionary guide...”

Let me save you some time: An AI agent is not magic.

It’s not “AGI” (Artificial General Intelligence). It’s not a some sort of a robot living in your cloud drive. And despite the hype, it’s not going to replace your entire department by Tuesday.

At its core, an AI agent is just three things: a brain, a set of tools, and a loop.

That’s it. Once you understand this simple recipe, 90% of the hype evaporates, and you’re left with something actually useful.

You're Already Using AI Agents (And Didn't Know It)

If you ask a modern AI like Claude, “What’s the weather in Amsterdam?” it will give you a real-time answer.

But here’s the secret: The AI doesn’t actually “know” the weather. That information wasn’t in its original training data. To give you that answer, the AI has to stop being a “chatbot” for a second and act like an agent.

Here is what’s happening under the hood:

The Prompt:

what is the current weather in Amsterdam? show me your thinking process

Answer:

Thinking process:

- Tool selection — Weather is real-time data that changes constantly, so I can't rely on my training knowledge. I used the weather_fetch tool with Amsterdam's coordinates (52.3676°N, 4.9041°E).
- Unit conversion — The API returned temperature in Fahrenheit (52.9°F). Since you're based in Amsterdam, I converted to Celsius: (52.9 - 32) × 5/9 ≈ 11.6°C.
- Data interpretation — The response included current conditions + a 5-day forecast. I surfaced the most relevant bits: current temp, sky condition, and notable changes ahead.

Result: It's currently sunny, ~11.6°C in Amsterdam. The week stays mild (around 13°C highs), with rain probability picking up on Wednesday (~40%).

The difference is this: A “Standard” AI is like a library. It’s a massive collection of books. It can tell you what’s inthe books, but it can’t leave the building.

• An AI Agent is the librarian who can actually walk out the front door, check the mail, make a phone call, and come back with an answer that isn’t in any of the books.

When you see an AI “browsing the web” or “checking a live price,” you aren’t just seeing a smarter chat, you are seeing an agent in action.

The Three Ingredients Every AI Agent Needs: Brain, Hands, Loop

Every agent, from a simple calendar bot to a complex coding assistant, is built from the same three pieces:

The Brain (The Language Model): Your Agent's Reasoning Engine

This is the “thinking” part (like Claude or Gemini). It reads your request and decides what to do. Its key skill is reasoning. You tell the AI, “You have access to these tools,” and it figures out which one to grab and when.

The Hands (The Tools): What Your Agent Can Actually Do

Without tools, an AI is just a brain in a jar. Tools are things the AI can actually do: search the web, read a spreadsheet, send an email, or check a database. The more tools you give it, the more “capable” it becomes.

The Loop (The Autonomy): What Makes It an Agent

This is what makes it an agent. The loop allows the AI to say: “Okay, I finished Step 1. Does the answer look right? No? Let me try Step 2.” It keeps going until the task is complete.

Brain + Hands + Loop = Agent. That’s the formula.

How to Build an AI Agent in Python (30 Lines of Code)

Enough theory. Let’s build a “Date & Deadline Agent.” It will be able to check today’s date and calculate exactly how many days are left until a specific deadline.

Even if you aren’t a developer, look at the code below. You’ll see it reads almost like English.

What you’ll need:

• Python installed.

• An Anthropic API key (from console.anthropic.com).

• In your terminal, run: pip install pydantic-ai

Create a file called my_agent.py to give the AI its instructions in VS Code and paste this in:

Do you want me to explain this part in more detail? Let me know in the comments!

Leave a comment

from pydantic_ai import Agent
from dotenv import load_dotenv
import datetime

load_dotenv() # Pulls your API key from a .env file

# 1. THE BRAIN: Give it a model and a personality
agent = Agent(
    model="anthropic:claude-sonnet-4-5",
    system_prompt="You are a helpful assistant. Use the available tools when you need current information."
)

# 2. THE HANDS: Define what the agent can actually DO
@agent.tool_plain
def get_current_time() -> str:
    """Returns the current date and time."""
    return datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")

@agent.tool_plain
def calculate_days_until(target_date: str) -> str:
    """Calculates days until a date (format: YYYY-MM-DD)."""
    target = datetime.datetime.strptime(target_date, "%Y-%m-%d")
    delta = target - datetime.datetime.now()
    return f"{delta.days} days until {target_date}"

# 3. THE RUN: Give it a task and let the loop take over
result = agent.run_sync(
    "What is today's date, and how many days until 2026-12-31?"
)

print(result.output)

When you run the code above, you will see a result like this:

Today's date is **March 8, 2026**, and there are **297 days** until December 31, 2026.

What just happened?

You didn't tell the agent which tool to use first. You just gave it a goal. The "Brain" saw your question, realized it didn't know the date, used the "get_current_time" tool, then used the "calculate_days_until" tool, and finally summarized the answer for you.

What AI Agents Can't Do: Cutting Through the Myths

Now that you’ve seen the “engine,” let’s kill some common myths:

• “Agents think for themselves.” Not really. They are pattern-matching engines with superpowers. They don’t have “desires”; they just follow the loop you set up.

• “Agents are dangerous.” An agent is only as powerful as the tools you give it. If you give it a calculator, it can do math. If you give it your credit card and a login to Amazon... well, that’s on you. You set the boundaries.

• “I need to be a genius to use them.” If you can follow a recipe, you can build an agent. The ecosystem is getting simpler every single day.

AI Agent vs. Claude: When Do You Actually Need One?

If Claude and Gemini can already check the weather, search the web, and run code, why do you need to know how to build your own?

Think of it like the difference between a Smartphone and a Smart Home System.

1. Use the “Built-in” AI (The Smartphone) when:

• You have a one-off question (”What’s the best flight to Tokyo?”).

• You need a quick summary or a brainstorm session.

• You’re using tools manually - you open the app, you take the photo, you send the text.

• The Vibe: “I just need a quick assist while I’m at the controls.”

2. Build a Custom Agent (The Smart Home) when:

• It needs to follow “Your Rules”: A smart home doesn’t just have a lightbulb; it knows to turn that light on only when you walk into the room after 6:00 PM. A custom agent follows your specific business logic.

• It needs to be “Always-On”: You want something running in the background - monitoring your emails, updating a spreadsheet, and pining your Slack - without you having to manually prompt it every ten minutes.

• It needs to connect “Hidden” Data: You want the AI to look at your private project files or your specific customer history, not just the general internet.

• The Vibe: “I want to build a system that works for me while I sleep.”

The Rule of Thumb: If you are the one typing every step and waiting for the answer, a regular AI chat is perfect. The moment you want the AI to orchestrate a process - connecting different apps and making decisions based on your specific “house rules”; that’s when you need a custom agent.

Which part clicked for you? Or which part still feels fuzzy? Drop it below; your question might be the next article!

Leave a comment

PS: If you're new here and wondering why a software engineer is writing about all this - here's why I started Becoming with AI."

Recently, I came across an app that helps non-English speakers build apps in minutes.
When I looked at their About section, one sentence stood out:

“We want to democratize app creation for people who don’t speak English.”

I genuinely liked that mission.
But it also made me think.

Because if that’s the goal, then, what exactly is broken in the current system, and what does “democratizing AI” actually mean?

How AI Tokenization Actually Works

Before we talk about fairness, access, or democratization, we need to talk about something much more basic: how large language models process text at all.

If you’ve ever worked with LLMs before, you probably know this:

AI doesn’t understand words.
It doesn’t understand characters.

It understands tokens.

And almost everything — pricing, limits, latency — is based on them.

Tokenizers break text into pieces that models can work with and tokenizers are not neutral.

Why English Is the Cheapest Language for AI

English has a “superpower”: it expresses meaning using very few tokens.

To make this concrete, I took the same content and wrote it in three languages that I am most familiar with:

• English

• Dutch

• Turkish

Then I ran them through the same tokenizer (https://platform.openai.com/tokenizer) .

Here’s what came out:

Same idea, same intent, same information. Very different token counts.

The result is actually not random:

• Dutch uses a lot of compound words, which tokenizers often split apart.

• Turkish is an agglutinative language, where meaning is built by stacking suffixes, and each stack adds more tokens.

• English, on the other hand, gets a smooth ride.

  Subscribe now

This Isn't a Bug. It's How Tokenizers Are Designed

It’s easy to call this unfair or broken. But technically speaking, nothing is “wrong“.

English dominates high-quality training data.
So tokenizers are optimized for English frequency patterns.

That means:

• common English words → single tokens

• common English suffixes → single tokens

• common English constructions → efficiently compressed

As a result:

English expresses meaning with fewer tokens than almost any other natural language.

The system is doing exactly what it was designed to do.

That means, English is the cheapest language to think in.

The Hidden Language Tax in AI Pricing

Most AI pricing today is token-based, directly or indirectly.

Which means non-English users hit limits faster, burn their credits sooner, or become “expensive” users without realizing why.

And none of this is visible at the UI level.

From the outside, everything looks inclusive.
From the inside, the system is quietly optimized around one language.

English dominates AI not only because of data, but because most large AI systems are built in the United States, where English is the default language for research, infrastructure, and early users.

That default shapes what gets optimized and what becomes expensive. Other languages work, but they are not the center of the system.

This is why it matters for other countries (especially non-English speaking ones) to build their own LLMs: not to replace global models, but to ensure their languages, costs, and ways of thinking are first-class citizens rather than afterthoughts.

If AI systems:

• think most cheaply in English

• price based on tokens

• and scale globally

Then we have to ask:

Who is actually paying the language tax?

Is it the user?
The company?
Or the people who never get to build at all?

Because real democratization will start when we design systems that don’t punish people for the language they think in.

And right now, we’re not quite there.

Many people’s first interaction with AI looks like this.
They open the chat, type something broad like “Help me think about my next article”: skim the response, and feel a quiet sense of dissatisfaction.

The output is fluent. Technically fine.
But nothing really moves.

I’ve had that reaction myself. Not because the AI failed, but because I hadn’t decided what I actually needed from it. The conclusion is usually the same: this is impressive, but not for me.

Subscribe now

Over time, I’ve noticed that clarity matters more than capability when it comes to using AI. To ask well, you first have to decide what kind of help you’re looking for: exploration or evaluation, support or resistance, speed or depth. That decision requires a certain kind of self-awareness, the ability to name:

• where you are

• and what you’re trying to reach

When people struggle to get value from AI, it’s rarely a tooling issue. It’s a hesitation to commit to a point of view. The sharper the intent, the more useful the response becomes. In that sense, AI doesn’t just help you think. It quietly reflects how much responsibility you’re willing to take for your own thinking.

I’ve found it helpful to distinguish between using AI as a tool and using it as a thinking partner.

When it’s treated as a tool, the interaction stays transactional. You ask, it answers, you move on.

When it’s treated as a thinking partner, something different happens. The exchange slows down. The responses push back a little. Weak framing becomes visible. Sometimes the reply isn’t helpful in itself, but it shows me what I was really trying to say, or what I was avoiding.

A thinking partner isn’t there to replace judgment. It engages with it. It surfaces assumptions, questions framing, and reflects reasoning back with less emotional attachment.

Used this way, AI becomes a kind of cognitive mirror. It doesn’t just respond to inputs; it reveals patterns in how you think. For leaders, writers, and people who build things, that distinction matters.

Tools optimize execution, thinking partners improve decision quality

Once that shift happens, the question changes. It’s no longer “What should I ask?” but “How do I invite better thinking?” and that’s where prompts start to matter.

A weak prompt sounds like this:

Help me write an article about AI and leadership.

It asks the AI to decide the angle, the depth, the audience, and the point of view, all things the writer should probably sit with first. The result is often polished and inoffensive, and easy to forget.

A stronger prompt takes responsibility before asking for help:

I’m writing for leaders who feel overwhelmed by AI hype. I believe AI is most useful as a thinking partner, not a productivity shortcut. Challenge this idea. Tell me where it’s naive. Help me sharpen it without softening it.

The difference isn’t the wording. It’s the posture.

One prompt hands thinking away.
The other invites collaboration.

Using AI this way also changes how it feels to create. The shift is subtle but important: from performing to exploring. When writing or making decisions in public-facing contexts, there’s often pressure to sound confident, decisive, and finished. That pressure leaks into how AI gets used. We ask for answers when what we really need is space.

Exploration asks for a different posture. It makes room for half-formed thoughts, contradictions, and early drafts that aren’t meant to impress anyone. AI lowers the cost of being unfinished, which makes it easier to stay with a question a little longer.

There’s something deeper underneath this that I don’t see talked about very often. When people create; write, decide, build, they’re not just assembling information. They’re bringing their values, experiences, doubts, and personal stakes into the work. That inner participation is what gives creation its weight.

You can call it judgment. Or intuition. Or, maybe even, soul.

Whatever you call it, it’s the part of thinking that can’t be automated. It’s what allows someone to stand behind an idea rather than simply present it. When AI is used to bypass that part, the output may still look fine, but it often feels strangely empty. Not wrong. Just… Ownerless.

When leaders and writers stop using AI to polish outcomes and start using it to explore ideas, the interaction becomes safer and more honest. In that space, clarity doesn’t come from sounding right. It comes from staying with a question long enough to understand it.

If AI reflects how you think, then the real question isn’t how good the tool is.
It’s how willing you are to stay present with unfinished thoughts.

The next time you open the chat, notice what you’re asking it to do:
think for you, or think with you.

I was listening to Tristan Harris on the Diary of a CEO podcast yesterday, and it sparked my curiosity about the AI race. He was talking about how companies are racing to build uncontrollable AI and what kind of future is ahead of us. I see that people are divided into two sides when it comes to the AI debate: those who see salvation and those who see extinction. Companies are investing a huge amount of money into AI to get ahead in this AI race, saying “if we don’t do it, they will” to rationalize their actions.

But what is this race all about? Finding solutions to every human problem, like curing cancer or making the world a better place? Or destroying the world by creating the most intelligent thing ever and not being able to control it at all? People are either on one side or the other. The answer, though, is not so simple.

When GPT-3 came out, it blew everybody’s mind about what AI could do. It helped with coding bugs, created fantastic reports for companies, and helped with school presentations. That was all great, but it didn’t show AI’s real potential, did it? Companies and governments started to discover that AI could do more than just help you write a 1000 word blog post. It can help detect cancer cells in mammogram results, help surveil citizens, or even kill people by targeting them automatically. It’s all so powerful, yet we’re not at AI’s full potential today.

The most important breakthrough hasn’t happened yet (Harris also mentioned this in the podcast): Recursive Self-Improvement. Imagine AI being aware of itself, recognizing its own weaknesses to complete a task at hand, and improving on those weaknesses. After this happens, which some experts expect that it could happen within 2 to 10 years, the world will change dramatically. Big companies leading the AI race, especially in the USA, believe this change is inevitable. On the other hand, critics say these AI labs should be regulated and companies should make sure AI is safe and won’t harm people. There’s an existential crisis for humanity in the shadow of AI.

Here, I raise two questions: Why would AI likely be more harmful than good? And is “controlling AI” even the real problem?

Subscribe now

Let’s look at the first question: If people are creating AI, why would it become potentially bad? Does that mean people are bad?

This question reminded me of the Sufi perspective on the creation of the world. They use a “mirror metaphor” to explain that creation is like a reflection of God. It appears inverted or reversed, which is why what is perfect and purely good in God can appear as imperfect or containing evil in the created world. I think what humans are trying to create with AI is similar to this analogy, with one key difference: God is perceived as perfect, but humans are not. We contain the capacity for both deep love and unspeakable cruelty.

Now we’re creating AI, a reflection of a reflection. Think about it: we’re creating something in our image, and we get to choose what to reflect. We can code in our problem solving abilities and our logic. But can we code in empathy? Grief? The ability to know that some things matter even when they don’t make logical sense?

The question isn’t whether AI will be “good” or “bad” but rather, what aspects of humanity are we coding into it, and what are we leaving out?

Now for the second question: Is “controlling AI” even the real problem? Let’s step back and think: why do we want to control AI? Because it can be harmful. But harmful to whom? Humans, right?

In the world we live in, we, humans, are the smartest of all creatures, at least until AI becomes self-aware and improves itself on its own. Then there are animals, which we sometimes use as a food source, make into shoes and coats, or love and adopt. We’re able to do that because we’re smarter than them. So what happens when something even smarter than us comes along? Do we become the “animals” of this world? Why would the world need humans anyway? I think this question is deeply shaking our most fundamental instinct: to be alive. No matter what, we want to live and we want to exist. That’s why some people see AI as an existential threat to humanity.

Humans may not be the smartest in the world anymore, but we have emotions, consciousness, and ethics that AI, we believe, can only mimic. I’m living in Amsterdam where the housing is a huge problem. I recently visited a cemetery with my boyfriend to see the graves of his grandparents and great-grandparents. If AI ran the city of Amsterdam, what would stop it from building houses on cemeteries and just burning all the dead bodies? From a pure logical perspective, this could make sense on paper while completely missing what makes Amsterdam Amsterdam: the layers of memory, grief, tradition, and meaning that can’t be quantified.

So where does this leave us?

This is the hardest part of the whole article. And honestly, I am not sure.

I believe in the world we’re living in, we need to hold onto our “human side” even more. Critical thinking and decision-making are more important than solving mathematical problems or knowing how to write an if statement in Python. We should be even more aware of our biases, the ones that make us blind to some facts or shape our thoughts about others.

I know we might not have time to figure this out before AI surpasses us. But I believe that a better future is possible with AI, only if we collectively work towards improving ourselves while the race is happening, not after it’s already over.

• Diary of a CEO: AI Expert: We have 2 years before Everything Changes! We Need To Start Protesting! — Tristan Harris

• https://www.bowdoin.edu/news/2025/10/keep-expressing-yourself-and-stay-human-urges-ai-expert-and-author-brian-christian.html

• The Alignment Problem: Machine Learning and Human Values