# Skill: Distinguish an AI agent from a chatbot and use the agent's action loop correctly

> Use this skill when explaining or applying the fundamental difference between a chatbot (talks) and an AI agent (acts) -- so users know what to expect, what to ask for, and when to reach for each.

Source: sodigi-learn -- vibecoding/kto-takoy-ii-agent - https://sodigi.io/learn/vibecoding/kto-takoy-ii-agent

## When to use
- A user is confused about why their AI tool is not "doing" anything -- just giving advice.
- Someone is trying to build something with a plain chat interface and getting frustrated by manual copy-pasting.
- You need to explain why an agent needs access to files, a terminal, or the internet.
- A user asks "how do I make AI actually build this for me, not just describe it?"

## Core rules
- A chatbot only produces text -- you do all the action yourself.
- An AI agent has tools (file access, shell commands, web search) and runs a loop: think -> act -> check result -> fix -> repeat.
- The underlying model can be identical -- what differs is whether it was given tools and permission to use them.
- Vibecoding means working with agents: you say WHAT, the agent figures out HOW and executes it.
- An agent's action loop is self-correcting: it runs code, sees the error, fixes it, re-runs -- without user intervention.
- Always tell the agent to act ("create," "run," "fix"), not to advise ("how do I," "explain how").
- A chatbot is the right tool for advice, explanations, and ideation. An agent is the right tool for execution.

## Procedure
1. Identify the user's actual need: advice/explanation (chatbot) vs. execution/output (agent).
2. If execution is needed, confirm the tool being used actually has agent capabilities (file access, terminal, web).
3. Write the prompt as a task with action verbs: "Create...", "Launch...", "Fix...", "Check..."
4. Grant the agent permission to act autonomously step by step, and instruct it to show results before continuing.
5. Let the agent run its think-act-check loop; intervene only when a step needs course-correction.

## Ready-to-use prompt
```
[Describe what you want built/done]. Work on your own, step by step:

1. First, ask me for any missing information: [what you need].
2. Create the necessary files yourself.
3. Run it and confirm it works without errors.
4. If something breaks -- fix it yourself and check again.
5. At the end, briefly tell me what you made and how to use it.
```

## Pitfalls
- Expecting a plain chat window to take action -- a chatbot only writes text; you still have to do everything else.
- Using "how do I..." when you mean "do it" -- advice prompts produce instructions, not results.
- Confusing a more expensive model with an agent -- smarter brain, still no hands without tools.
- Fearing the agent will "do everything wrong" -- it shows every step; you can stop and redirect at any moment.
- Asking a chatbot to create files or run commands -- it cannot; you need a tool with actual agent capabilities.