👤

The Vibe Architects

Corwin VibeCoding for Beginners

Od 0 do 1 u pravljenju AI Agenata

The DevBreak Engineers

Tvoj napredak kroz kurs 0%

🔒 Moraš prvo da položiš test prethodnog modula.

MODUL 1

Uvod u Vibe Coding

Upoznaj se sa konceptom, razumi zašto menja način na koji se gradi softver, i pronađi pravi alat za sebe.

Šta je Vibe Coding?

Termin "vibe coding" prvi je lansirao Andrej Karpathy (bivši direktor AI-ja u Tesli, jedan od suosnivača OpenAI-ja) početkom 2025. godine. Ideja je jednostavna: umesto da pišeš kod red po red, ti opisuješ šta želiš, a AI to pretvara u kod.

        Karpathyjev originalni opis:
        Postoji nova vrsta programiranja koju zovem "vibe coding", gde se potpuno prepustiš vibracijama, prihvataš eksponencijalni rast, i zaboraviš da kod uopšte postoji. Moguće je jer su LLM-ovi postali toliko dobri.
      

U 2026. godini, ovo više nije novitet. Prema nedavnim istraživanjima, preko 90% američkih developera svakodnevno koristi AI alate za kodiranje, a oko 41% svog koda u svetu je AI-generisano. Collins Dictionary je čak proglasio "vibe coding" za reč 2025. godine.

Tri glavne komponente Vibe Coding-a

Mindset + metoda: Gradiš softver intuitivno, sa minimumom trenja.
Flow state: Mentalno stanje u kome ideje dolaze lako, a progres deluje prirodno.
Alati: AI asistenti (Claude, GPT, Gemini), live coding okruženja (Replit, Lovable, Bolt), i laki jezici (Python, JavaScript).

Zašto Vibe Coding menja igru

🔧 Tradicionalno programiranje

Pisanje koda red po red
Manuelno debagovanje
Potrebno duboko znanje sintakse
Spori iterativni ciklusi
Visoka početna barijera

✨ Vibe Coding

Prompt + fino podešavanje AI output-a
AI pomaže u debagovanju i ispravkama
Traži dobro formulisan problem
Brze build–test–iterate petlje
Niska barijera za ulazak

Kome Vibe Coding najviše koristi?

Startap osnivači - brzo grade prototipove da testiraju ideje.
Istraživači - validiraju hipoteze u satima umesto nedeljama.
Developeri - automatizuju repetitivne zadatke pomoću malih agenata.
Ne-programeri - prvi put mogu da grade funkcionalan softver.

        💡 Zanimljivo za 2026:
        Prema najnovijim podacima, 63% aktivnih vibe coding korisnika nije iz programerskih pozadina. Product menadžeri, dizajneri i osnivači prave cele aplikacije opisujući ih prirodnim jezikom.
      

Šta su AI Agenti?

AI agent je pametan digitalni radnik. Napaja ga Large Language Model (LLM), ima pristup alatima, memoriji i logici. Može da rasuđuje, koristi alate i pamti zadatke.

Primeri AI agenata iz stvarnog sveta:

Agent koji skreira stranice i sumira dnevne promene.
Chatbot koji obrađuje korisničke žalbe, automatski ih klasifikuje i piše odgovore.
Generator ideja za društvene mreže na osnovu prethodnih postova.
Lični asistent koji zakazuje sastanke i piše mejlove.

Alati za Vibe Coding u 2026.

Pejzaž alata se drastično promenio od 2024. godine. Ovde je ažurirana lista najpopularnijih alata, podeljena po nivou tehničkog znanja:

🚀 Za ne-programere i osnivače (App Builderi)

Alat	Najbolje za	Cena (2026)
Lovable	Kompletna full-stack web apps iz razgovora	od $25/mes
Replit (sa Agent-om)	Cloud IDE + deploy u jednom mestu	od $20/mes
Bolt.new	Brzi full-stack prototipovi	od $20/mes
v0 (Vercel)	UI komponente i frontend	od $20/mes
Base44	Početnicima prijatna alternativa	od $20/mes

⚙️ Za developere (AI Code Editori)

Alat	Snaga	Cena (2026)
Cursor	Najdublje razumevanje velikih codebase-ova	od $20/mes
Claude Code	Terminal agent, autoriše ~4% svih javnih GitHub commit-ova	Uključeno u Claude planove
Windsurf	Cascade feature, pristupačniji od Cursor-a	od $15/mes
GitHub Copilot	Najjeftinija opcija sa unlimited completions	od $10/mes

        ⚠️ Upozorenje za troškove:
        Iako mnogi alati imaju free tier, aktivni builderi u 2026. godini obično troše $50–$150 mesečno preko svih platformi. Replit posebno može neočekivano skočiti zbog usage-based credits. Uvek postavi spending limits.
      

What is Vibe Coding?

The term "vibe coding" was launched by Andrej Karpathy (former Tesla AI Director, OpenAI co-founder) in early 2025. The idea is simple: instead of writing code line by line, you describe what you want and AI turns it into code.

        Karpathy's original description:
        There is a new kind of coding I call "vibe coding", where you fully give in to the vibes, embrace exponential growth, and forget that code even exists. It's possible because LLMs have gotten so good.
      

In 2026, this is no longer a novelty. According to recent research, over 90% of American developers use AI coding tools daily, and around 41% of all code worldwide is AI-generated. Collins Dictionary even named "vibe coding" the word of 2025.

Three Main Components of Vibe Coding

Mindset + method: Build software intuitively, with minimum friction.
Flow state: A mental state where ideas come easily and progress feels natural.
Tools: AI assistants (Claude, GPT, Gemini), live coding environments (Replit, Lovable, Bolt), and lightweight languages (Python, JavaScript).

Why Vibe Coding Changes the Game

🔧 Traditional Programming

Writing code line by line
Manual debugging
Deep syntax knowledge required
Slow iterative cycles
High entry barrier

✨ Vibe Coding

Prompt + fine-tuning of AI output
AI helps with debugging and fixes
Requires a well-formulated problem
Fast build–test–iterate loops
Low entry barrier

Who Benefits Most from Vibe Coding?

Startup founders — quickly build prototypes to test ideas.
Researchers — validate hypotheses in hours instead of weeks.
Developers — automate repetitive tasks with small agents.
Non-programmers — build functional software for the first time.

        💡 Interesting for 2026:
        According to the latest data, 63% of active vibe coding users have no programming background. Product managers, designers and founders are building entire applications by describing them in natural language.
      

What are AI Agents?

An AI agent is a smart digital worker. Powered by a Large Language Model (LLM), it has access to tools, memory and logic. It can reason, use tools and remember tasks.

Real-world AI Agent Examples:

An agent that scrapes web pages and summarizes daily changes.
A chatbot that handles customer complaints, auto-classifies them and writes responses.
A social media idea generator based on previous posts.
A personal assistant that schedules meetings and writes emails.

Vibe Coding Tools in 2026

The tool landscape has changed dramatically since 2024. Here is the updated list of the most popular tools, divided by technical skill level:

🚀 For Non-Programmers and Founders (App Builders)

Tool	Best For	Price (2026)
Lovable	Complete full-stack web apps from conversation	from $25/mo
Replit (with Agent)	Cloud IDE + deploy in one place	from $20/mo
Bolt.new	Fast full-stack prototypes	from $20/mo
v0 (Vercel)	UI components and frontend	from $20/mo
Base44	Beginner-friendly alternative	from $20/mo

⚙️ For Developers (AI Code Editors)

Tool	Strength	Price (2026)
Cursor	Deepest understanding of large codebases	from $20/mo
Claude Code	Terminal agent, authors ~4% of all public GitHub commits	Included in Claude plans
Windsurf	Cascade feature, more affordable than Cursor	from $15/mo
GitHub Copilot	Most affordable with unlimited completions	from $10/mo

        ⚠️ Cost Warning:
        Although many tools have a free tier, active builders in 2026 typically spend $50–$150 per month across all platforms. Replit in particular can unexpectedly spike due to usage-based credits. Always set spending limits.
      

🎯 Spreman si za proveru znanja?

Završio si čitanje ovog modula. Sada je vreme da testiraš šta si naučio.
Potrebno je 50%+ na bodovanim pitanjima da bi prešao na sledeći modul.

📝 Provera znanja — Modul 1

Odgovori na pitanja kako bi utemeljio svoje znanje

💡 Napomena o bodovanju: Uvodna pitanja (označena kao "Nebodovano") služe samo da ti pomognu da se setiš osnova — ne utiču na finalni rezultat. Bodovana pitanja određuju da li možeš da pređeš na sledeći modul.

NEBODOVANO uvodna provera

1. Ko je prvi popularizovao termin "vibe coding"?

Sam Altman

Andrej Karpathy

Satya Nadella

Dario Amodei

BODOVANO

2. Koja od navedenih NIJE glavna razlika između tradicionalnog programiranja i vibe coding-a?

Brzina iteracije

Niža barijera za ulazak

Korišćenje računara

AI pomaže u debagovanju

BODOVANO

3. Koji alat je najprikladniji za ne-programera koji prvi put pravi full-stack aplikaciju?

GitHub Codespaces

Lovable ili Replit

Zed

Vim sa Copilot pluginom

BODOVANO

4. AI agent se razlikuje od običnog chatbot-a zato što:

Generiše samo tekst

Može da odgovori brže

Može da koristi alate, pamti i rasuđuje kroz više koraka

Uvek je jeftiniji od chatbot-a

OTVORENO PITANJE

5. Opiši svojim rečima u 2-3 rečenice: šta bi mogao da napraviš za 1 sat pomoću vibe coding-a, što ti je ranije trebalo nedeljama?

PRAKTIČNA VEŽBA

6. Otvori Lovable (ili Bolt.new) i napiši jedan prompt koji opisuje jednostavnu aplikaciju koju bi hteo da napraviš. Prilepi prompt u polje ispod:

Tvoj rezultat

MODUL 2

Principi Vibe Coding-a

Pet temeljnih principa koji razlikuju uspešan projekat od frustrirajuće bele magle. Plus: kako napraviti PRD koji AI stvarno razume.

Pet osnovnih principa

Thinking

→

Framework

→

Checkpoints

→

Debugging

→

Context

Thinking: Kako razbiti problem na korake.
Framework: Gradivni blokovi na kojima stojiš.
Checkpoints: Pauze da testiraš progres i razmisliš.
Debugging: Pronalaženje i popravka problema.
Context: Okolne informacije koje tvoj kod treba.

1. Thinking — Kako misliti kao programer

Svako dobro vibe coding iskustvo počinje jasnim razmišljanjem. Četiri nivoa razmišljanja koji su ti potrebni:

Piramida razmišljanja

Procedural: Celokupna slika (Dinner Party)
Computational: Tehnike (tehnike kuvanja)
Analytical: Recepti (strukturirana rešenja)
Logical: Sastojci (osnovna logika)

Analogija sa kuvanjem

Ako gradiš aplikaciju za porudžbinu hrane, moraš razmišljati na svim nivoima — od osnovne logike (šta je jelo?), preko analitičkog (kako se spaja sa korpom?), pa do proceduralnog (kako izgleda celo iskustvo korisnika?).

2. Framework — Frontend, Backend, API

Razmisli o aplikaciji kao o restoranu:

🍽️ Frontend (Restoran & konobar)

Ono što korisnik vidi: dugmad, tekst, slike, dizajn.

Napravljen sa: HTML, CSS, JavaScript, React, Vue, Angular

👨‍🍳 Backend (Kuhinja & kuvar)

Obrađuje zahteve, čuva podatke, izvodi logiku.

Napravljen sa: Python, Node.js, baze podataka

        🔗 API
        API je dostavna služba — ono što povezuje tvoj restoran sa drugim restoranima. Kada tvoja aplikacija šalje poruku, koristi email servis kroz API. Kada plaća, koristi Stripe API.
      

Zašto su frameworks važni u vibe coding-u?

Brža gradnja — ne izmišljaš točak svaki put.
Lakše debagovanje — problemi su dokumentovani.
Više fokusa na ideji, manje na kodu.

        💡 Savet:
        Ne prihvataj slepo što AI predlaže. Pitaj ga: "Zašto si izabrao Next.js umesto Astro?" — i naučićeš ekosistem dok gradiš.
      

3, 4, 5. Checkpoints, Debugging, Context

Checkpoints — Posle svake velike promene, proveri da li aplikacija radi. Nemoj graditi 10 fitčera pa onda testirati.

Debugging — Kada AI zaglavi, dodaj što više konteksta: error poruka, screenshot, opis šta si očekivao.

Context — AI radi mnogo bolje kada mu daš relevantne fajlove, šemu baze, postojeći kod. Ne očekuj od njega da pročita tvoje misli.

PRD — Product Requirements Document

PRD je dokument koji opisuje šta treba da se napravi. Ključan je za vibe coding jer AI-ju daje jasne smernice.

Šta PRD treba da sadrži:

Overview — šta aplikacija radi i za koga?
Key Features — glavne funkcionalnosti
Required Skills/Tech — koje tehnologije
User Stories — scenariji iz života korisnika

        💡 Savet:
        Pusti svoj omiljeni chatbot (Claude, ChatGPT, Gemini) da ti pomogne da napraviš PRD. Zatim ga koristi kao kontekst kada kreneš da gradiš u Lovable, Cursor ili Replit.
      

Deployment — Od lokalnog do živog

Deployment znači objavljivanje aplikacije na internetu da bi drugi ljudi mogli da je koriste. Misli o tome kao o prelasku iz "draft" u "live".

Dok ne uradiš deploy, aplikacija radi samo na tvom računaru.

Šta se dešava tokom deploy-a:

Frontend se hostuje (da korisnici vide sajt)
Backend radi na serveru (da odgovara na zahteve)

        Moderni hosting u 2026:
        Vercel (za frontend i Next.js), Netlify (jednostavan static hosting), Railway i Render (backend + baze), Replit (sve u jednom). Mnogi vibe coding alati imaju ugrađen deploy na jedan klik.
      

VEŽBA

Napravi svoj PRD

Koraci:

Izaberi ideju za jednostavnu mobilnu ili web aplikaciju.
Koristi chatbot (Claude, ChatGPT, Gemini) da ti pomogne da napraviš PRD posebno za vibe coding.
Primeni PRD u nekom alatu (Lovable, Bolt, Replit).

The Five Core Principles

Thinking

→

Framework

→

Checkpoints

→

Debugging

→

Context

Thinking: How to break a problem into steps.
Framework: The building blocks you stand on.
Checkpoints: Pauses to test progress and reflect.
Debugging: Finding and fixing problems.
Context: The surrounding information your code needs.

1. Thinking — How to Think Like a Programmer

Every good vibe coding experience starts with clear thinking. Four levels of thinking you need:

Thinking Pyramid

Procedural: The big picture (Dinner Party)
Computational: Techniques (cooking techniques)
Analytical: Recipes (structured solutions)
Logical: Ingredients (basic logic)

Cooking Analogy

If you're building a food ordering app, you must think on all levels — from basic logic (what is a dish?), through analytical (how does it connect to the cart?), to procedural (what does the entire user experience look like?).

2. Framework — Frontend, Backend, API

Think of your application as a restaurant:

🍽️ Frontend (Restaurant & Waiter)

What the user sees: buttons, text, images, design.

Built with: HTML, CSS, JavaScript, React, Vue, Angular

👨‍🍳 Backend (Kitchen & Chef)

Processes requests, stores data, runs logic.

Built with: Python, Node.js, databases

        🔗 API
        API is the delivery service — what connects your restaurant to other restaurants. When your app sends a message, it uses an email service through an API. When it accepts payment, it uses the Stripe API.
      

Why are Frameworks Important in Vibe Coding?

Faster building — you don't reinvent the wheel every time.
Easier debugging — problems are documented.
More focus on the idea, less on the code.

        💡 Tip:
        Don't blindly accept what AI suggests. Ask it: "Why did you choose Next.js instead of Astro?" — and you'll learn the ecosystem while building.
      

3, 4, 5. Checkpoints, Debugging, Context

Checkpoints — After every major change, verify the app works. Don't build 10 features and then test.

Debugging — When AI gets stuck, add as much context as possible: error message, screenshot, description of what you expected.

Context — AI works much better when you give it relevant files, database schema, existing code. Don't expect it to read your mind.

PRD — Product Requirements Document

A PRD is a document that describes what needs to be built. It's crucial for vibe coding because it gives AI clear guidelines.

What a PRD Should Contain:

Overview — what the app does and for whom?
Key Features — main functionality
Required Skills/Tech — which technologies
User Stories — real-life user scenarios

        💡 Tip:
        Let your favorite chatbot (Claude, ChatGPT, Gemini) help you create a PRD. Then use it as context when you start building in Lovable, Cursor or Replit.
      

Deployment — From Local to Live

Deployment means publishing your application on the internet so others can use it. Think of it as going from "draft" to "live".

Until you deploy, your app only runs on your computer.

What Happens During Deployment:

Frontend is hosted (so users can see the site)
Backend runs on a server (to respond to requests)

        Modern Hosting in 2026:
        Vercel (for frontend and Next.js), Netlify (simple static hosting), Railway and Render (backend + databases), Replit (all-in-one). Many vibe coding tools have built-in one-click deploy.
      

EXERCISE

Create Your Own PRD

Steps:

Choose an idea for a simple mobile or web app.
Use a chatbot (Claude, ChatGPT, Gemini) to help you create a PRD specifically for vibe coding.
Apply the PRD in a tool (Lovable, Bolt, Replit).

🎯 Vreme za test!

Prošao si kroz sve lekcije ovog modula. Hajde da proverimo šta si naučio.

📝 Provera znanja — Modul 2

BODOVANO

1. Koja su pet principa Vibe Coding-a?

Planning, Coding, Testing, Deployment, Marketing

Thinking, Framework, Checkpoints, Debugging, Context

Design, Develop, Deploy, Debug, Document

Ideas, Tools, Code, Review, Launch

BODOVANO

2. U analogiji restorana, šta predstavlja API?

Meni restorana

Kuhinju

Dostavnu službu koja povezuje restorane

Recepte

BODOVANO

3. Šta je PRD?

Product Requirements Document - dokument koji opisuje šta aplikacija treba da radi

Programming Rules Document

Python Runtime Developer

Project Review Deployment

BODOVANO

4. Zašto su checkpoints važni u vibe coding-u?

Jer AI zahteva periodne pauze

Jer je bolje testirati inkrementalno nego graditi 10 fitčera pa onda debagovati sve odjednom

Jer je to pravilo Git-a

Nisu važni

OTVORENO PITANJE

5. Objasni svojim rečima razliku između Frontend-a i Backend-a koristeći primer iz stvarnog sveta (ne restoran).

PRAKTIČNA VEŽBA

6. Koristeći chatbot, napravi mini-PRD za aplikaciju po svom izboru. Prilepi PRD u polje ispod (samo glavne tačke: Overview, Features, Users).

Tvoj rezultat

MODUL 3

Osnove AI i LLM-ova

Demistifikuj AI pojmove. Razumi šta je LLM, razliku između LLM-a i ChatGPT-ja, temperature, multi-modalnost i izbor provajdera.

AI → ML → DL → GenAI

Pre nego što razumeš šta su AI agenti, moraš da razumeš hijerarhiju pojmova.

Artificial Intelligence
Tehnika koja omogućava mašinama da imitiraju ljudsko ponašanje

Machine Learning
Statističke metode za učenje iz iskustva

Deep Learning
Višeslojne neuronske mreže

Generative AI
Stvara novi tekst, slike, video, audio

Kratka istorija Generative AI-ja

1930-1950: Alan Turing i njegove inovacije (Turingova mašina, Turingov test)
1986: Yann LeCun revolucionizuje neuronske mreže kroz backpropagation
1990-e: Machine Learning se razvija
2006: Deep Learning se reintrodukuje
2014: GAN-ovi (Generative Adversarial Networks) prave slike
2017: Transformers se uvode ("Attention is All You Need")
2020: GPT-3 revolucionizuje AI generisanje teksta
2022: ChatGPT se lansira, AI postaje mainstream
2023-2024: GPT-4, Claude 3, Gemini — multimodalni modeli
2025-2026: Era agenta — Claude 4.x, GPT-5.x, agentic workflows

Šta su Large Language Models (LLM)?

LLM = Large Language Model. Raščlanjeno:

Large — veliki trening podaci i velika veličina modela (milijarde parametara)
Language — fokus na obrade prirodnog jezika
Model — pravi predikcije bazirane na naučenim šablonima

        Kako LLM radi?
        LLM je generativni model koji dopunjava tekst sa najverovatnijom sledećom rečju. To su transformer modeli koji rade tako što obraćaju pažnju na važne reči u rečenici.
      

Primer: Kada LLM dobije input "Glavni grad Francuske je", on generiše verovatnoće za sledeću reč:

Paris    → 90%
dense    → 2.4%
diverse  → 1.2%
home     → 0.8%
a        → 0.2%
      

LLM vs ChatGPT — Koja je razlika?

Mnogi ljudi koriste "LLM" i "ChatGPT" kao sinonime. Nisu. Evo pravilne razlike (ažurirano za 2026.):

🧠 LLM (Mozak)

Trenirana na ogromnim količinama teksta
Predviđa sledeću reč
Generalna namena
Primeri (2026): GPT-5.5, Claude Opus 4.7, Gemini 2.5, LLaMA 3, Mistral, DeepSeek
Kao motor auta — moćan, ali sirov

🤖 ChatGPT / Claude.ai / Gemini (Proizvod)

Proizvod napravljen oko LLM-a
Ima chat interfejs (memorija razgovora)
Ima sistemske instrukcije (pravila ponašanja)
Ima sigurnosne slojeve
Ima alate (web pretraga, code execution, image generation)

Ključni parametri LLM-a

🌡️ Temperature

Temperature je parametar koji određuje koliko je model "kreativan":

Niska (0-0.3): Konzistentni, predvidivi odgovori. Dobro za faktičke zadatke, kod.
Srednja (0.4-0.7): Balans između kreativnosti i preciznosti. Dobro za većinu zadataka.
Visoka (0.8-1.5): Kreativni, raznovrsni odgovori. Dobro za brainstorming, kreativno pisanje.

👁️ Multi-Modal LLM-ovi

Moderni LLM-ovi (svi top modeli u 2026.) rade sa više vrsta ulaza: tekst, slike, audio, video, kod.

LLM Provajderi u 2026.

💳 Plaćeni API provajderi (gotovi, najjači modeli)

Provajder	Najpoznatiji modeli (2026)	Specijalnost
Anthropic	Claude Opus 4.7, Claude Sonnet 4.6, Claude Haiku 4.5	Kodiranje, reasoning, agenti
OpenAI	GPT-5.5, GPT-5.5 Pro, GPT-5.4	Opšta inteligencija, multimodalnost
Google	Gemini 2.5 Pro, Gemini 2.5 Flash	Dugačak kontekst, multimodalnost
DeepSeek	DeepSeek R1, DeepSeek V3	Reasoning, pristupačne cene

🆓 Open-Source provajderi

Hugging Face: Ogromna biblioteka modela, zahteva moćnu mašinu
Ollama: Optimizovano lokalno pokretanje, manja biblioteka
Together AI, Groq, Fireworks: Hosted open-source modeli

AI → ML → DL → GenAI

Before you understand AI agents, you need to understand the hierarchy of terms.

Artificial Intelligence
Technique enabling machines to imitate human behavior

Machine Learning
Statistical methods for learning from experience

Deep Learning
Multi-layer neural networks

Generative AI
Creates new text, images, video, audio

Brief History of Generative AI

1930–1950: Alan Turing and his innovations (Turing machine, Turing test)
1986: Yann LeCun revolutionizes neural networks through backpropagation
1990s: Machine Learning develops
2006: Deep Learning is reintroduced
2014: GANs (Generative Adversarial Networks) create images
2017: Transformers introduced ("Attention is All You Need")
2020: GPT-3 revolutionizes AI text generation
2022: ChatGPT launches, AI becomes mainstream
2023–2024: GPT-4, Claude 3, Gemini — multimodal models
2025–2026: The agent era — Claude 4.x, GPT-5.x, agentic workflows

What are Large Language Models (LLMs)?

LLM = Large Language Model. Broken down:

Large — huge training data and large model size (billions of parameters)
Language — focus on natural language processing
Model — makes predictions based on learned patterns

        How LLMs Work?
        An LLM is a generative model that completes text with the most likely next word. These are transformer models that work by paying attention to important words in the sentence.
      

Example: When an LLM receives "The capital of France is", it generates probabilities for the next word:

Paris    → 90%
dense    → 2.4%
diverse  → 1.2%
home     → 0.8%
a        → 0.2%

LLM vs ChatGPT — What's the Difference?

Many people use "LLM" and "ChatGPT" as synonyms. They're not. Here is the proper difference (updated for 2026):

🧠 LLM (The Brain)

Trained on enormous amounts of text
Predicts the next word
General purpose
Examples (2026): GPT-5.5, Claude Opus 4.7, Gemini 2.5, LLaMA 3, DeepSeek
Like a car engine — powerful, but raw

🤖 ChatGPT / Claude.ai / Gemini (The Product)

Product built around an LLM
Has a chat interface (conversation memory)
Has system instructions (behavior rules)
Has safety layers
Has tools (web search, code execution, image generation)

Key LLM Parameters

🌡️ Temperature

Temperature determines how "creative" the model is:

Low (0–0.3): Consistent, predictable responses. Good for factual tasks and code.
Medium (0.4–0.7): Balance between creativity and precision. Good for most tasks.
High (0.8–1.5): Creative, diverse responses. Good for brainstorming and creative writing.

👁️ Multi-Modal LLMs

Modern LLMs (all top models in 2026) work with multiple input types: text, images, audio, video, code.

LLM Providers in 2026

💳 Paid API Providers (ready-to-use, strongest models)

Provider	Top Models (2026)	Specialty
Anthropic	Claude Opus 4.7, Sonnet 4.6, Haiku 4.5	Coding, reasoning, agents
OpenAI	GPT-5.5, GPT-5.5 Pro, GPT-5.4	General intelligence, multimodality
Google	Gemini 2.5 Pro, Gemini 2.5 Flash	Long context, multimodality
DeepSeek	DeepSeek R1, DeepSeek V3	Reasoning, affordable pricing

🆓 Open-Source Providers

Hugging Face: Huge model library, requires powerful hardware
Ollama: Optimized local running, smaller library
Together AI, Groq, Fireworks: Hosted open-source models

🎯 Vreme za test!

Završio si Modul 3. Hajde da proverimo koliko si savladao AI osnove.

📝 Provera znanja — Modul 3

BODOVANO

1. Koja je tačna hijerarhija pojmova od najšireg ka najužem?

ML → AI → DL → GenAI

GenAI → AI → ML → DL

AI → ML → DL → GenAI

DL → ML → GenAI → AI

BODOVANO

2. Šta radi LLM u svojoj srži?

Pretražuje internet u realnom vremenu

Dopunjava tekst predviđajući najverovatniju sledeću reč

Traži odgovore u bazi podataka

Koristi Google pretragu ispod haube

BODOVANO

3. Koja je prava razlika između LLM-a i ChatGPT-ja?

Nema razlike, sinonimi su

LLM je noviji tip modela

LLM je "mozak" (sirov model), ChatGPT/Claude.ai su proizvodi izgrađeni oko LLM-a sa interfejsom, memorijom, alatima i sigurnosnim slojevima

ChatGPT koristi isključivo Anthropic tehnologiju

BODOVANO

4. Koju temperaturu bi postavio ako želiš kreativne, raznovrsne odgovore (npr. brainstorming)?

Visoku (0.8-1.5)

Nultu (0)

Negativnu (-0.5)

Srednju (0.5)

BODOVANO

5. Šta znači "multi-modalni" LLM u 2026.?

LLM koji ima više različitih verzija

LLM koji može da obrađuje više tipova ulaza: tekst, slike, audio, video, kod

LLM koji može da priča na više jezika

LLM koji ima više različitih temperatura

OTVORENO PITANJE

6. Kada bi gradio AI agenta za ozbiljan projekat, koji LLM provajder bi izabrao i zašto? (Navedi najmanje 2 razloga)

PRAKTIČNA VEŽBA

7. Idi u Anthropic Console ili Google AI Studio. Testiraj isti prompt sa dve različite system poruke (npr. "sarkastični" vs "formalni"). Kratko opiši razliku u odgovorima.

Tvoj rezultat

MODUL 4

AI Agenti — Arhitektura

Razumi šta čini agenta: memorija, specijalizacija, RAG i alati. Ovo su gradivni blokovi svakog ozbiljnog AI projekta.

Od LLM-a do AI Agenta

Sirovi LLM dobija tekst i vraća tekst. AI agent radi mnogo više:

        Definicija AI Agenta:
        AI agent je sistem koji koristi LLM kao "mozak", ali ima dodatne sposobnosti: pristup alatima, memoriju, bazu znanja, i sposobnost da planira i izvršava više koraka prema zadatom cilju.
      

Memorija Agenta

LLM po sebi odgovara samo na trenutni input (to je text completion model). Da bi agent "razgovarao" i "pamtio", treba mu memorija.

Tip	Analogija	Zadatak
Kratkoročna memorija	Sticky notes na stolu	Traje dok traje trenutni razgovor
Dugoročna memorija	Sveska starih projekata	Čuva stare razgovore i dohvata relevantne poruke kad treba
Refleksije	Rezime učenja nakon pregleda beleški	Procesuiraju memoriju i prave rezimee, zaključke, činjenice

Specijalizacija LLM-a

Opšti LLM je dobar u mnogo stvari, ali nije ekspert ni u jednoj po difoltu. Dva načina specijalizacije:

🎓 Fine-tuning

Trajni pristup — menja model
Treba stotine ili hiljade primera
Treba veliku računarsku snagu
Za stabilne, ne-menjajuće domene

📚 RAG (Retrieval Augmented Generation)

Brz, ne-trajni pristup
LLM "gleda" u specijalnu biblioteku pre nego što odgovori
Radi i sa malim ili promenljivim podacima
Lakše za implementaciju

        💡 Praktičan savet:
        U 99% slučajeva za vibe coding projekte, počni sa RAG-om. Fine-tuning zahteva ML ekspertizu i veliki trud.
      

RAG — Retrieval Augmented Generation

        Analogija studenta na ispitu:
        LLM je kao pametan student na ispitu. Njegovo znanje je ono što je naučio pre ispita (trening data). RAG mu daje pristup cheat sheet-u (retrieval sistem) sa najnovijim ili domen-specifičnim znanjem.
      

Kako RAG radi — korak po korak:

Pretvaraš dokumente u "vektore" koristeći embedding model
Čuvaš u vector database-u (Pinecone, Chroma, Weaviate, Qdrant)
Kada postaviš pitanje, sistem traži najsličnije značenje teksta (cosine similarity)
LLM kombinuje dobijen tekst sa sopstvenim sposobnostima → tačan odgovor

Alati — Šta agent zapravo radi

LLM-ovi su sjajni u generisanju teksta, ali alati su im potrebni da:

Računaju numeričke formule (calculator tool)
Pretražuju web (search tool)
Šalju email (email tool)
Pristupaju bazi podataka
Izvršavaju kod (code execution)

        Standard 2026: MCP (Model Context Protocol)
        Anthropic je pokrenuo MCP — otvoreni standard za povezivanje AI agenata sa alatima i podacima. Sada praktično svi glavni provajderi podržavaju MCP.
      

From LLM to AI Agent

A raw LLM receives text and returns text. An AI agent does much more:

        AI Agent Definition:
        An AI agent is a system that uses an LLM as its "brain", but has additional capabilities: access to tools, memory, a knowledge base, and the ability to plan and execute multiple steps toward a given goal.
      

Agent Memory

An LLM by itself only responds to the current input (it's a text completion model). For an agent to "converse" and "remember", it needs memory.

Type	Analogy	Role
Short-term memory	Sticky notes on the desk	Lasts for the current conversation
Long-term memory	Notebook of old projects	Stores old conversations and retrieves relevant ones when needed
Reflections	Summary of learning after reviewing notes	Processes memory and creates summaries, conclusions, facts

LLM Specialization

A general LLM is good at many things, but is not an expert in any by default. Two ways to specialize:

🎓 Fine-tuning

Permanent approach — changes the model
Requires hundreds or thousands of examples
Requires large computational power
For stable, non-changing domains

📚 RAG (Retrieval Augmented Generation)

Fast, non-permanent approach
LLM "looks" into a special library before responding
Works with small or changing data
Easier to implement

        💡 Practical tip:
        In 99% of vibe coding projects, start with RAG. Fine-tuning requires ML expertise and significant effort.
      

RAG — Retrieval Augmented Generation

        Student Exam Analogy:
        An LLM is like a smart student at an exam. Their knowledge is what they learned before the exam (training data). RAG gives them access to a cheat sheet (retrieval system) with the latest or domain-specific knowledge.
      

How RAG Works — Step by Step:

Convert documents into "vectors" using an embedding model
Store in a vector database (Pinecone, Chroma, Weaviate, Qdrant)
When you ask a question, the system finds the most similar meaning (cosine similarity)
LLM combines the retrieved text with its own capabilities → accurate answer

Tools — What an Agent Actually Does

LLMs are great at generating text, but they need tools to:

Calculate numerical formulas (calculator tool)
Search the web (search tool)
Send emails (email tool)
Access a database
Execute code (code execution)

        2026 Standard: MCP (Model Context Protocol)
        Anthropic launched MCP — an open standard for connecting AI agents with tools and data. Now virtually all major providers support MCP.
      

🎯 Vreme za test!

Sada razumeš arhitekturu AI agenata. Hajde da proverimo!

📝 Provera znanja — Modul 4

BODOVANO

1. Koja je ključna razlika između sirovog LLM-a i AI agenta?

LLM je brži

AI agent ima memoriju, pristup alatima i bazi znanja, plus sposobnost planiranja više koraka

Nema razlike

Agent može da zaboravi, LLM ne može

BODOVANO

2. Šta su "refleksije" u kontekstu memorije agenta?

Trenutni razgovor

Database tabele

Procesirana memorija koja pravi rezimee, zaključke i činjenice iz starih razgovora

Tip prompt-a

BODOVANO

3. Kada bi izabrao fine-tuning umesto RAG-a?

Uvek — fine-tuning je bolji

Kada imaš stabilni domen, hiljade primera, ML ekspertizu, i trebaš da model "govori" tvoj domen nativno

Kada imaš male i promenljive podatke

Kada želiš brzu implementaciju

BODOVANO

4. Šta je vektorska baza podataka u kontekstu RAG-a?

Obična SQL baza

Tabela sa tekstom

Pretraživa memorija koja čuva dokumente kao numeričke "vektore" (brojeve koji predstavljaju značenje)

Backup sistem

BODOVANO

5. Zašto agent treba alate ako LLM već generiše tekst?

Jer LLM ne može pouzdano da računa matematiku, pretražuje web u realnom vremenu, šalje mejlove ili pristupa bazama — alati mu omogućavaju da zaista utiče na svet

Jer su alati marketinški zahtev

Jer su jeftiniji

Nije potrebno, LLM može sve sam

OTVORENO PITANJE

6. Opiši AI agenta za customer support za online prodavnicu. Koje alate bi mu dao? Koju memoriju? Kakav RAG?

PRAKTIČNA VEŽBA

7. Zamisli da gradiš AI helper za određeni use-case (npr. student life, travel guide, study guide). Odgovori: Koje dokumente bi indeksirao? Koja pitanja bi AI trebalo da zna da odgovori? Zašto je RAG bolji od fine-tuning-a ovde?

Tvoj rezultat

MODUL 5

Od 0 do 1 — AI Agenti u Praksi

Vreme je da gradiš. Biranje LLM-a, RAG u praksi, indeksiranje dokumenata, GIGO princip i sklapanje punog agenta.

Kako birati LLM za svog agenta

LLM je kičma svakog agenta. Izbor ogromno utiče na kvalitet, cenu i brzinu.

🦙 Ollama (lokalno)

Ograničena biblioteka modela
Optimizovano korišćenje resursa uređaja
Besplatno, bez API troškova
Privatnost — podaci ne napuštaju tvoju mašinu

🤗 Hugging Face

Ogromna biblioteka modela
Veći modeli traže visok RAM i GPU
Inference API dostupan online
Idealno za specijalizovane modele

        💡 Moja preporuka za 2026:
        Za vibe coding & agente: Claude Sonnet 4.6 ili Opus 4.7
Za troškovno-svesne projekte: Gemini 2.5 Flash ili DeepSeek R1
Za lokalno/privatno: LLaMA 3.x ili Gemma 3 preko Ollama
Za multimodalne use-case-ove: GPT-5.5 ili Gemini 2.5 Pro

      

Indeksiranje dokumenata — kritične odluke

Retrieval koristi cosine similarity, pa dužina chunk-ova teksta direktno utiče na tačnost retrieval-a.

📦 Chunks su preveliki

Nerelevantna informacija se meša u rezultate
Teže je naći precizan match
Troše se tokeni bezveze

🧩 Chunks su premali

AI može propustiti kontekst
Razbija se logička celina
Dobija se fragmentiran odgovor

Pravilo za 2026: Standardni chunk size je 200-800 tokena sa preklapanjem 10-20%. Ali moraš testirati za svoj konkretan slučaj.

        ⚠️ GIGO princip — Garbage In, Garbage Out
        Kada imaš puno podataka, čišćenje postaje kritično. Loše pripremljeni podaci → loši odgovori, bez obzira koliko je LLM pametan.
      

Sklapanje punog agenta — 5 koraka

Definiši cilj agenta — šta tačno treba da radi? Za koga?
Biraj LLM — uzevši u obzir cenu, kvalitet, multimodalnost
Dizajniraj memoriju — kratkoročna, dugoročna, refleksije
Dodaj RAG sloj — ako je potreban domen-specifičan kontekst
Opremi alatima — web search, database, calculator, email, itd.

Popularni frameworks za agente u 2026:

LangGraph — standardni framework za graf-based agente
CrewAI — multi-agent sisteme
Claude Code SDK — za agente bazirane na Claude
OpenAI Agents SDK — zvanični OpenAI stack
Vercel AI SDK — za TypeScript/JavaScript agentima

How to Choose an LLM for Your Agent

The LLM is the backbone of every agent. The choice hugely impacts quality, cost and speed.

🦙 Ollama (local)

Limited model library
Optimized device resource usage
Free, no API costs
Privacy — data never leaves your machine

🤗 Hugging Face

Massive model library
Larger models require high RAM and GPU
Inference API available online
Ideal for specialized models

        💡 My recommendation for 2026:
        For vibe coding & agents: Claude Sonnet 4.6 or Opus 4.7
For cost-conscious projects: Gemini 2.5 Flash or DeepSeek R1
For local/private: LLaMA 3.x or Gemma 3 via Ollama
For multimodal use cases: GPT-5.5 or Gemini 2.5 Pro

      

Document Indexing — Critical Decisions

Retrieval uses cosine similarity, so chunk length directly impacts retrieval accuracy.

📦 Chunks Are Too Large

Irrelevant information mixes into results
Harder to find a precise match
Tokens are wasted

🧩 Chunks Are Too Small

AI may miss context
Logical units are broken up
Fragmented answers are returned

Rule for 2026: Standard chunk size is 200–800 tokens with 10–20% overlap. But you must test for your specific case.

        ⚠️ GIGO Principle — Garbage In, Garbage Out
        When you have a lot of data, cleaning becomes critical. Poorly prepared data → poor answers, no matter how smart the LLM is.
      

Building a Full Agent — 5 Steps

Define the agent's goal — what exactly should it do? For whom?
Choose an LLM — considering cost, quality, multimodality
Design memory — short-term, long-term, reflections
Add a RAG layer — if domain-specific context is needed
Equip with tools — web search, database, calculator, email, etc.

Popular Agent Frameworks in 2026:

LangGraph — standard framework for graph-based agents
CrewAI — multi-agent systems
Claude Code SDK — for agents based on Claude
OpenAI Agents SDK — official OpenAI stack
Vercel AI SDK — for TypeScript/JavaScript agents

🎯 FINALNI TEST!

Stigao si do kraja! Ovo je poslednji test. Odluči tvoju dalju putanju.

🏁 Finalni test — Modul 5

Poslednja provera pre sertifikacije

BODOVANO

1. Kada bi izabrao Ollama umesto plaćenog API-ja?

Nikada, API je uvek bolji

Samo za enterprise klijente

Kada trebaš privatnost, lokalno procesiranje, ili bez API troškova — idealno za eksperimente i sensitive data

Samo za Windows

BODOVANO

2. Šta je GIGO princip?

Generic Input, Generic Output

Garbage In, Garbage Out - loš kvalitet unetih podataka = loši rezultati

Google Input, Google Output

Giga In, Giga Out

BODOVANO

3. Šta se dešava kada su chunks preveliki u RAG sistemu?

Sistem radi brže

Nerelevantna info se meša u rezultate i teže je naći precizan match

RAG postaje skuplji

Ništa

BODOVANO

4. Kojih pet koraka je za sklapanje punog AI agenta?

Kodiranje, testiranje, deploy, marketing, skaliranje

Design, Develop, Deploy, Debug, Document

Definicija cilja → Izbor LLM-a → Dizajn memorije → RAG sloj → Dodavanje alata

Write code, test, launch

BODOVANO

5. Koji framework je standardni za graf-based AI agente u 2026?

LangGraph

TensorFlow

React

Laravel

OTVORENO PITANJE

6. Opiši prednosti RAG-a u odnosu na fine-tuning za startup koji pravi chatbot za svoju dokumentaciju. (Daj najmanje 3 razloga)

FINALNA PRAKTIČNA VEŽBA

7. Napravi kompletan plan AI agenta. Popuni sve tačke:

📌 Use-case:
📚 Dokumenti koje bi indeksirao:
❓ Tipovi pitanja koje agent treba da odgovori:
🤖 LLM koji bi izabrao (i zašto):
🔧 Alate koje bi dodao:
💾 Gde bi ga hostovao:

Tvoj finalni rezultat

🏆

Corwin VibeCoding — Advanced

You understand AI agents.
Now it's time to build them.

You just proved you have the foundation. The next step isn't another theory course — it's hands-on, no-code agent building that gets you to real results in real time.

🎙️ WEBINAR EXCLUSIVE — Expires May 3

$630 $87

Save $543 today — 86% off. No code required.

--Days

--Hrs

--Min

--Sec

You've come this far. The edge belongs to those who act on what they know.

⏳ The Window Is Closing

The AI builder economy is not waiting for anyone to feel ready. The gap between those who understand AI and those who deploy it is where all the value is being created — right now.

97M

new AI-related jobs created by 2026 (World Economic Forum)

$150k+

avg. consulting revenue earned by top 1% of no-code AI builders in 2025

40%

productivity increase in teams that deploy custom AI workflows

Understanding the theory is the entry ticket. Building with it is the advantage.

🗺️ What You'll Actually Build

8 hands-on modules. No Python. No terminal. Just you, your tools, and real agents that work.

The No-Code Agent Stack

Your toolkit: n8n, Make.com, Dify — and how they map to every concept from this course

Your First Automated AI Workflow

Build a content pipeline that researches, writes, and publishes — fully automated

RAG Without Code — AI That Knows Your Data

Upload your documents and build a chatbot that answers questions about your business, visually

Multi-Agent Teams

Orchestrate specialized agents — a researcher, a writer, and a reviewer — working together

Real-World Integrations

Connect your agents to Slack, Gmail, Notion, and CRMs — where real business lives

Memory, Context & Personalization

Build agents that remember users, adapt over time, and feel genuinely intelligent

Deployment, Monitoring & Monetization

Ship your agent publicly, track performance, and learn how to charge clients for it

Capstone — Build & Launch a Real AI Product

End-to-end project: define, build, deploy, and present a real agent to a real audience

Everything in this course is designed for exactly where you are right now — one step past the theory.

or keep reading

📦 Everything Included

This isn't a video library you'll never finish. It's a structured sprint from concept to deployment.

8 no-code project modules with step-by-step walkthroughs
Workflow templates you can deploy immediately
Private community of builders at the same level
Live Q&A sessions with The DevBreak Engineers
Lifetime access — including all future updates
Certificate of completion for your portfolio

⚠️ After May 3, the price increases to $126.

The $87 webinar rate is a one-time arrangement tied exclusively to this week's event. Once it expires, the course moves to $126 — which is still an 80% discount from the standard enrollment price. We cannot guarantee that a promotion of this scale will return, or when. If you've been considering it, the time to act is now.

One Decision. Real Leverage.

🎙️ WEBINAR EXCLUSIVE — Expires May 3

$630 $87

One-time payment. Lifetime access.

🔒 If you go through the first two modules and feel it's not right for you, email us within 7 days for a full refund. No questions asked.

💪

Odlično si prošao kurs!

Uspešno si završio sve module i savladao osnove vibe coding-a i AI agenata. Za naš Advanced kurs potrebno je 80%+ na finalnom testu — još malo fali!

        Preporuka:
        Možeš ponovo da pokušaš testove gde si imao manje od 80%. Takođe, pogledaj besplatne resurse na početku kursa da utvrdiš znanje.
      

Build AI agents. No code required.

Get started

What would you like to improve or achieve using AI?

Before You Begin

⚙️ The Tactical Setup

🗺️ The Roadmap: What to Expect

Uvod u Vibe Coding

Šta je Vibe Coding?

Tri glavne komponente Vibe Coding-a

Zašto Vibe Coding menja igru

🔧 Tradicionalno programiranje

✨ Vibe Coding

Kome Vibe Coding najviše koristi?

Šta su AI Agenti?

Primeri AI agenata iz stvarnog sveta:

Alati za Vibe Coding u 2026.

🚀 Za ne-programere i osnivače (App Builderi)

⚙️ Za developere (AI Code Editori)

What is Vibe Coding?

Three Main Components of Vibe Coding

Why Vibe Coding Changes the Game

🔧 Traditional Programming

✨ Vibe Coding

Who Benefits Most from Vibe Coding?

What are AI Agents?

Real-world AI Agent Examples:

Vibe Coding Tools in 2026

🚀 For Non-Programmers and Founders (App Builders)

⚙️ For Developers (AI Code Editors)

🎯 Spreman si za proveru znanja?

📝 Provera znanja — Modul 1

Principi Vibe Coding-a

Pet osnovnih principa

1. Thinking — Kako misliti kao programer

Piramida razmišljanja

Analogija sa kuvanjem

2. Framework — Frontend, Backend, API

🍽️ Frontend (Restoran & konobar)

👨‍🍳 Backend (Kuhinja & kuvar)

Zašto su frameworks važni u vibe coding-u?

3, 4, 5. Checkpoints, Debugging, Context

PRD — Product Requirements Document

Šta PRD treba da sadrži:

Deployment — Od lokalnog do živog

Šta se dešava tokom deploy-a:

Napravi svoj PRD

The Five Core Principles

1. Thinking — How to Think Like a Programmer

Thinking Pyramid

Cooking Analogy

2. Framework — Frontend, Backend, API

🍽️ Frontend (Restaurant & Waiter)

👨‍🍳 Backend (Kitchen & Chef)

Why are Frameworks Important in Vibe Coding?

3, 4, 5. Checkpoints, Debugging, Context

PRD — Product Requirements Document

What a PRD Should Contain:

Deployment — From Local to Live

What Happens During Deployment:

Create Your Own PRD

🎯 Vreme za test!

📝 Provera znanja — Modul 2

Osnove AI i LLM-ova

AI → ML → DL → GenAI

Kratka istorija Generative AI-ja

Šta su Large Language Models (LLM)?

LLM vs ChatGPT — Koja je razlika?

🧠 LLM (Mozak)

🤖 ChatGPT / Claude.ai / Gemini (Proizvod)

Ključni parametri LLM-a

🌡️ Temperature

👁️ Multi-Modal LLM-ovi

LLM Provajderi u 2026.

💳 Plaćeni API provajderi (gotovi, najjači modeli)

🆓 Open-Source provajderi

AI → ML → DL → GenAI

Brief History of Generative AI

What are Large Language Models (LLMs)?

LLM vs ChatGPT — What's the Difference?

🧠 LLM (The Brain)

Build AI agents.
No code required.

You understand AI agents.
Now it's time to build them.