Every AI company now sells at least four words for “type something and get something back.” Chat. Project. Skill. Agent. The AI agent vs chatbot question is mostly marketing running faster than the people it is marketed to.
Here is the plain version. Use the simplest option that gets the job done. For most people on most days, that is a plain chat. The other three solve specific problems, and you only need one of them once you hit the problem it solves.
Below: what each term means, how to tell which one fits your task, and a prompt that helps you decide.
Quick answer: start with a plain chat. Move to a project when you keep retyping the same background, to a skill when you keep retyping the same steps, and to an agent only when the AI has to choose its own next steps using connected tools. Fancier is something you earn, not where you start.
Want the prompt first? Jump to the prompt.
The four terms in plain English
Names and features differ from tool to tool, and what you get depends on your plan. The ideas hold up anyway.
- Chat. One conversation. You ask, it answers, you correct it. Best for a one-time question, a draft, or an explanation.
- Project. A workspace that keeps related chats, files, and instructions together so you stop re-explaining the background. OpenAI describes a ChatGPT project as a dedicated space that can hold chats, files, instructions, and related context. Anthropic’s help center pairs projects with work that builds up context over time.
- Skill. A saved set of instructions for a job you do the same way every time. Anthropic’s help center describes skills as the tool for repeatable procedures you want applied automatically, and projects as the tool for accumulated context. If you want help writing the instructions themselves, start with Better AI Instructions: Stop Chasing Better Prompts. If you saved Gems in Gemini, see Google Is Replacing Gemini Gems. Save Your Instructions Before November.
- Agent. An AI that decides its own next steps and uses tools to carry them out, such as email, a calendar, files, or a web browser. Anthropic’s engineers draw the line like this. In a workflow, fixed code sets the steps. In an agent, the AI directs its own process and tool use. OpenAI’s guide for businesses describes agents in similar terms, as systems that carry out tasks for you with some independence.
These are not rival products. You can combine them. A project can hold the chats where you use a skill.
| Term | Use it when | Main thing to check |
|---|---|---|
| Chat | You have a one-time question, draft, or explanation | Facts and figures in the answer |
| Project | You keep retyping the same background | Which file an answer actually relied on |
| Skill | You keep retyping the same steps | That the output follows the steps every time |
| Agent | The job has many steps, information in several places, and decisions along the way | The actual record or sent folder, not the “done” message |
AI agent vs chatbot: which one do you need?
- A one-time question, draft, or explanation: chat.
- Ongoing work where you keep retyping the same background: project.
- A job you do the same way every time, where you keep retyping the same steps: skill, or saved instructions if your tool calls them something else.
- A job with many steps, information in several places, and decisions along the way, where the AI has to choose what to do next and use connected tools: agent.
- Not sure which problem you have: chat. You can always move up later.
Anthropic’s engineers give the same advice to the people building these systems: find the simplest solution possible, and add complexity only when you need it. If that is the rule for the people building agents, it is probably fine for the rest of us.
Try this: let the AI sort it out
Copy this into whichever AI tool you already use. Fill in the brackets with something non-confidential.
I do this task regularly: [describe the task in one or two sentences]. How often: [daily / weekly / monthly / once in a while]. What I usually give the AI: [notes, a spreadsheet, a pasted email, nothing]. Where the information lives: [email, calendar, shared files, my head]. Tell me whether a plain chat is enough, or whether I would get something real from a project, saved instructions (a skill), or an agent that can use connected tools. Start with the simplest option that works. If you recommend something more complicated, tell me what I would have to set up, what it could see or change, and where I should check its work. Do not assume I have any particular tool, plan, or feature.
Then check the recommendation against what your tool actually offers. The AI may not know which features are on your account unless you tell it.
A realistic example
Say you handle the weekly update for a volunteer committee. You might fill in the prompt like this: the task is turning your meeting notes into a short email to members, it happens weekly, you give the AI your typed notes, and the information lives in your notes and your email drafts.
A useful answer should sound something like this (the wording will vary): a plain chat is enough to start, so paste the notes and ask for the email. If you find yourself retyping the same instructions every week, save them as a skill or saved instructions. An agent would only make sense if you wanted it to send the email itself, which means it could see and change your email, and for a weekly note that is probably more than the job needs.
If the answer is vague, reply with this: “Pick one of the four. What is the first thing I should do today, and what should I check afterward?”
Where this goes wrong
- Assuming fancier is better. An agent on a simple writing task adds setup and more things to check, and it gives you nothing a chat would not.
- Assuming stored means used. A file sitting in a project does not mean every answer drew on it. If it matters, ask which file the answer relied on, then open that file and check.
- Assuming connected means unlimited. When you connect an app, the AI gets some access, not necessarily all of it. Read what you are granting before you click, and keep the sensitive material out, as covered in What Not to Put Into ChatGPT or Another AI Tool.
- Believing “done.” An agent that says it updated a record or sent a message is telling you what it did, or what it meant to do. Check the record. Check the sent folder. Polished confidence is not proof, as explained in Why AI Sounds So Confident When It’s Wrong.
The short version
Start with a chat. If you keep retyping the same background, make it a project. If you keep retyping the same steps, save a skill. If you want an agent, be able to say what it may see and change before you switch it on.
And if a vendor tells you that you need all four by Friday, that is a sales call, not a requirement.
