Agents! Thinking! Skills! Subagents! Cowork!
AI products have become more sophisticated. Some might say more intelligent. We have a come a lot of ways from prompting to a chat-bot. Or have we?
All of these ‘features’ are marketing speak for ‘prompting a Large Language Model‘, combined with conventional software. Sometimes not even that. Segmenting all of these things into different functions is quite useful for technical users. At the cost of alienating average users.
In this post, I will try to demystify these features. By explaining how they are just prompts.
AI Chat is a prompt…
…that is very very long.
Back in the days of GPT2(then the mOsT dAnGeRoUs model ever!!![1]), OpenAI launched a text completion tool. You gave it a prompt and it continued the text. Like autocorrect on your phone, but using a better technology. So if a model can extend any text, it can predict the next part of a conversation. You can learn more about how that works(in simple terms), in a previous post I made.
AI Chat is you, the user, prompting the model with the whole text of the chat, every time you respond. Whether using ChatGPT, Claude, Google Gemini, or any other model. And every response from the model is a continuation of that long prompt.
AI Thinking is a prompt…
…with a lot of steps.
We have established that Large Language Models(LLMs) don’t “know” anything. They just show you the most likely combination of words that might come up next. Which means they generate a lot of mistakes. If fact, if you were to prompt GPT5.5 or Claude Opus 4.7 without any instructions to maintain accuracy, you will get functional gibberish.
Which is why all AI chat products have a system prompt[2]. But that is not enough. So researchers created thinking or reasoning models. These models are trained to generate steps or chain-of-thought before their final output. These thinking steps become part of your chat. And what is an LLM chat?
Skills are prompts…
…that live in a file.
These models are trained on a lot of data. Petabytes. From classic novels, to washing machine repair manuals, to reddit comments. Which means they are quite generalized. They can’t give you an expert response on anything. Which is fine if you are looking to generate a 5 paragraph essay to cheat on your English homework. Not fine when you need it to process last week’s sales data and create a report.
Anthropic, makers of Claude, understood this problem. So they came up with the Skills framework. A way to turn the model into more of a specialist. How, you ask? why, what else but a prompt!
A prompt that lives in a file called SKILL.md that you can reuse for specific purposes. It is like having a specialist at your beck and call, if that specialist gives you different solutions to the same problem twice.
Agents are prompts…
…that run in the background.
Agents, or subagents, are AI chats that you can run in the background. How are they different than skills? They can integrate multiple skills and MCPs(see below) to complete the task. They are like any chat session or response running parallel to your main response. This allows you to run multiple tasks at the same time.
MCPs are prompts…
… with lots of technical info.
The issue early in the AI boom post-ChatGPT launch was that AI could not speak with other conventional computing resources. Apps and data sources, whether local, remote or on the web, do not speak human languages like english. They have APIs. OpenAI’s solution to this was “function calling”[3]. They let the model respond in a machine readable text that triggers some code.
But it was cumbersome. You had manually define each function or tool that you want to use, and write programs to go along with them.
Anthropic, makers of Claude, came up with the Model Context Protocol[4]. In simple terms, it is a bunch of definition of actions your AI app can use to interact with a other apps, services, commands, etc. While the server/client infrastructure is a lot of code, the underlying definitions are in plain text. Which means its a prompt.
All the way down
What we’ve learned today is that there more than one way to skin a cat prompt a LLM. Whether using skills, MCPs, or subagents. However, if they are not discrete implementations of specific features, why are they branded as such?
There are two ways to look at this:
- These firms want to present their products as more sophisticated than they are, or
- They want users to use their products in a specific way, and are branding for that purpose.
Either way, it does introduce confusion. I hope I have been able to clear some of that.
References
[1] https://slate.com/technology/2019/02/openai-gpt2-text-generating-algorithm-ai-dangerous.html
[3] https://developers.openai.com/api/docs/guides/function-calling
[4] https://modelcontextprotocol.io/docs/getting-started/intro
I do not have analytics on this site. So leave a comment, even a nasty one, here or on your preferred social site, or shoot an email to feedback at the domain of this site. I also have weekly newsletter where I share wisdom I have acquired in the form of quotes and notes. You can check the archive here. You can sign up below to receive it every week.