The development field has always evolved rapidly, but nothing compares to the era of artificial intelligence we are living in now. It’s becoming hard to stay up to date with so many new things to learn: agents, subagents, MCP, skills. How and when do I use all of this?
I was having this difficulty. I didn’t know how and where to start, I bought courses, watched videos, read articles but what really taught me was getting hands-on. I paid for the Claude Code subscription and started using AI to program.
A lot of things started to make sense to me, and I’d like to share what I’m learning on my journey to stay up to date in the era of artificial intelligence here on my blog. And before anything else, I think it’s important to explain the most common concepts and terms related to AI development.
Table of Contents
LLM (Large Language Model)
The LLM is the "big brain" behind AI. It is trained on a massive amount of data, allowing it to process information and return a result.
Each trained model has its own name and version, such as Claude 2.0, GPT-4, Claude Opus 4.6, Gemini 3.1 Pro, etc. Each has its own characteristics, capabilities, and limitations. Some are better at understanding conversational context, while others are faster or have a larger token limit.
When we use an AI service like ChatGPT, Claude, or Gemini, we are actually using a user-friendly interface that interacts with an LLM behind the scenes. The LLM is what truly processes the information and generates the response, while the interface makes it accessible and easy for us to use.
Although we often treat everything as "AI," it’s important to differentiate:
- LLMs → focused on text
- Image Models → image generation (Text-to-Image)
- Audio Models → voice synthesis (TTS)
- Multimodal Models → combine text, image, audio, etc.
It is also possible to chain several LLMs to create a more complex system where each model is responsible for a specific part of the process. For example, one LLM can be used to understand the context of a question, while another is specialized in generating the response.
Text -> LLM (interpretation) -> LLM (generation) -> TTS (Audio Model) -> Audio
Tools like LangChain help structure this type of workflow.
Tokens
Tokens are the basic units AI uses to read and respond. Instead of whole letters or words, the AI breaks text into chunks (like syllables or parts of words) through a process called tokenization.
For example, the word "house" might be split into tokens, and each token is converted into numbers so the AI can understand, process, and respond.
house -> [hou, se] -> [123, 456]
Most LLMs have a token limit they can process at once. If the prompt or response exceeds this limit, the AI may fail to process everything, leading to incomplete answers or errors. Paid plans often allow you to increase this limit.
Prompt
A prompt is the instruction, question, input, or command you send to the AI. This step is crucial because the quality of the response depends heavily on how the prompt is formulated. If you want a shorter, clearer, and more objective answer, it’s important to be specific in your prompt.
It is also vital to provide any relevant information so the AI understands the context and can respond appropriately.
Example: "Give me a list of 5 Brazilian cities with over 1 million inhabitants, formatted in JSON, following this example: { 'city': 'São Paulo', 'population': 12300000 }."
Response:
[
{ "city": "São Paulo", "population": 12300000 },
{ "city": "Rio de Janeiro", "population": 6748000 },
{ "city": "Brasília", "population": 3055000 },
{ "city": "Salvador", "population": 2887000 },
{ "city": "Fortaleza", "population": 2687000 }
]
Providing adequate context is key. Recently, I asked an AI to investigate a problem in my application; it started reading thousands of files in the node_modules folder, took several minutes, and consumed all my available tokens. When my tokens refreshed, I asked it to investigate the problem while specifically consulting the documentation of the tool I was using. Result: it solved the problem in a few seconds.
Embeddings
Embeddings are numerical representations of the meaning of words or texts. This allows the AI to understand that "house" and "home" are similar concepts, even if they are written differently. Words with similar meanings (like "dog" and "pet") stay numerically close, helping the AI understand context within a sentence.
An AI doesn't "know" everything before responding; it identifies patterns and predicts the next word based on what has been said and what it learned during training.
If a question is about a subject it doesn't know, it might invent an answer (called "hallucination"). To avoid this, techniques like RAG can be used to provide real information and ensure more accurate answers.
RAG (Retrieval-Augmented Generation)
As explained in LLM, AI is trained on a vast amount of data, but this knowledge has an expiration date. It's as if the LLM is frozen in time. It doesn't have access to information that happened after data collection or data that wasn't present during training.
To solve this, the RAG technique is used. It allows the AI to consult external sources — like your documents or company manuals—before responding. This prevents it from making things up and ensures responses are based on real facts.
MCP (Model Context Protocol)
Just as RAG is a technique to improve accuracy, MCP is a protocol that standardizes how AI connects to and interacts with different data sources and apps.
It acts as a bridge between the AI and the outside world, allowing it to access real-time information like database records, web search results, or even interact with other apps, such as a calendar to schedule a meeting or an email system to send a message.
Check out the list of popular MCP servers for Claude.
AI Agent
In a standard conversation, the AI receives a question (input) and returns an answer (output). But it can't do much beyond that. An AI Agent, however, is a more advanced system capable of performing specific tasks autonomously.
For example, if you ask an AI agent to "Schedule a meeting with John tomorrow at 10 AM," it can access your calendar, check John's availability, send an invite, and confirm the meeting—all without you doing anything else.
Examples of AI Agents: Claude Code, OpenCode, Copilot, AutoGPT, BabyAGI, AgentGPT, etc.
Skills
Skills are reusable instructions or capabilities we give to an agent so it becomes an expert at a certain task. We can specify how and when an agent should search the web, perform math, generate images, or follow a specific pattern.
.skills/
/summarize-text/
SKILL.md
Skill format (skill.md):
---
name: summarize-text
description: Summarize any given text in a clear and concise way. Use when the user asks for summaries or when dealing with long content.
---
# Summarize Text Skill
## Instructions
- Always generate a concise summary
- Keep the main ideas only
- Avoid unnecessary details
- Use bullet points when appropriate
- Maintain the original meaning
## Output format
- Short paragraph OR
- Bullet points (if content is complex)
Subagents
Like skills, subagents are instructions to be followed when performing a task via the main agent. By default, a main agent handles small tasks one at a time. With subagents, it can delegate these tasks to be executed in parallel or combined.
Subagent Example
---
name: RestaurantsFinder
description: Look for specific restaurants considering reviews and popularity given a location.
---
---
name: PlacesCloseToMe
description: Return only a list of places near a location and a range. Return the place name, popularity, and distance.
---
Command:
Give me a list of 10 Japanese restaurants near me. I am located at Praça da Sé in São Paulo. The maximum distance is 30 kilometers. Use the RestaurantsFinder and PlacesCloseToMe subagents.
Answer:
Here are the 10 Japanese restaurants closest to you at Praça da Sé, within 30km:
┌─────┬────────────────────────┬──────────────────────────┬───────────┐
│ # │ Restaurants │ Rating │ Distance │
├─────┼────────────────────────┼──────────────────────────┼───────────┤
│ 1 │ Izakaya Issa │ 4.5/5 │ ~1.5 km │
├─────┼────────────────────────┼──────────────────────────┼───────────┤
│ 2 │ Lamen Kazu │ 4.4/5 (1.043 reviews) │ ~1.8 km │
├─────┼────────────────────────┼──────────────────────────┼───────────┤
│ 3 │ Momo Lamen House │ 4.3/5 (518 reviews) │ ~1.8 km │
├─────┼────────────────────────┼──────────────────────────┼───────────┤
│ 4 │ Mugui │ 4.5/5 (294 reviews) │ ~1.9 km │
├─────┼────────────────────────┼──────────────────────────┼───────────┤
│ 5 │ Restaurante Porque Sim │ 4.5/5 (701 reviews) │ ~1.9 km │
├─────┼────────────────────────┼──────────────────────────┼───────────┤
│ 6 │ Sushi Kenzo │ 4.4/5 │ ~1.9 km │
├─────┼────────────────────────┼──────────────────────────┼───────────┤
│ 7 │ Sushi Isao │ 4.5/5 (649 reviews) │ ~2.0 km │
├─────┼────────────────────────┼──────────────────────────┼───────────┤
│ 8 │ Tanka Restaurante │ 4.5/5 (1.171 reviews) │ ~2.0 km │
├─────┼────────────────────────┼──────────────────────────┼───────────┤
│ 9 │ Jojo Ramen │ 4.4/5 │ ~3.5 km │
├─────┼────────────────────────┼──────────────────────────┼───────────┤
│ 10 │ Djapa Moema │ 4.8/5 (7.212 reviews) │ ~8.0 km │
└─────┴────────────────────────┴──────────────────────────┴───────────┘
The first 8 are located in the Liberdade neighborhood (SP's Japanese district), just a few minutes' walk away. The highlight by rating is Djapa Moema (4.8/5 with over 7,000 reviews), despite being the furthest on the list.
SDD (Spec Driven Development)
SDD is a document containing all the details, rules, and specifications of your project. It is usually the first thing done before starting. Much like the Prompt, it’s important to spend time here to guide the AI toward the expected result. In an SDD, you specify the motivation, goal, technologies, business rules, and entities. You can even use AI to help write a good Spec, using tools like kiro.dev.
Conclusion
AI is not the future; it is already among us today. There is a fear it might take our jobs, and there are reasons for that concern. But at the same time, AI can be a great ally in our daily lives—automating tasks, scouring files, and acting as a powerful pair programming partner.