Around the Model
Everything built around the weights
An AI you talk to today is not just a model. It is a model in the middle of a set of rings: instructions, context, tools, a loop, connectors, skills, memory, and finally you. People talk about MCP and skills as if they were new kinds of intelligence. They aren't. Each one is another ring, built by people, around a centre that doesn't change.
Scroll to add the rings one at a time, and watch the same question, "What's on my plate tomorrow?", get a better answer each time. On a wide screen you can fold the text away and let the rings fill the page.
A model, alone
At the centre is a set of frozen weights: billions of numbers set during training, then locked. Give it text and it predicts what text comes next. That is the whole capability.
It cannot look anything up. It does not know today's date. It has no idea who you are. Ask it about your day and it will do what it always does: produce the most likely-sounding continuation.
Instructions
The first ring is a paragraph of text, the system prompt, placed before every conversation. Who to be, what to refuse, how to admit uncertainty.
It changes behaviour without changing a single weight. That idea is the key to every ring that follows.
Context
The context window is everything the model can see right now: the conversation so far, a pasted document, a file, today's date. It has no other way to see.
Paste in your schedule and the model can sort it, spot the clash, suggest what to move. But someone had to go and fetch it, and that someone was you.
Tools
Here the model gets hands. Alongside your question it receives a list of tools, each with a name, a description and the arguments it takes. Instead of answering, the model can write a request to use one.
The model still touches nothing. It writes a request; the program around it (the harness) decides whether to run it, runs it, and puts the result back into context.
The loop
One tool call is a reflex. Do it again and again (read the result, decide the next step, act) and it becomes an agent. Try one thing, notice it failed, try another, then answer.
The loop is still outside the model. The harness keeps calling it until it says it is done.
MCP: one socket for everything
Every app used to connect every tool its own way: a hundred apps and a hundred services meant ten thousand custom connections. The Model Context Protocol, proposed by Anthropic in 2024, is a shared shape for the connection. A service builds one MCP server, and any AI that speaks MCP can plug into it. People call it the USB-C of AI.
A calendar, a notes app, a code host or a news archive each builds its connector once, and it works in Claude, ChatGPT, Gemini and the rest. The model doesn't get smarter. It gets somewhere to look.
Skills: know-how, on demand
MCP gives the model reach. A skill gives it know-how: a folder holding a SKILL.md of instructions, plus any scripts or templates it needs.
The trick is what the model sees and when. At first it sees only each skill's one-line description. It reads the full instructions only when the task calls for them, so a hundred skills cost almost nothing until one is needed. On the ring, the dashed cards are skills that are present but not yet opened.
docx, pdf, pptx and xlsx are skills Anthropic publishes.Memory
Close the window and the model forgets everything. Always. What we call memory is saved notes, often plain files like MEMORY.md, that the harness reads back into context at the start of the next session.
The model doesn't remember you. The system around it keeps a note, and hands the note back.
You
Every ring was put there by a person. You choose which tools it may run, which connectors to trust, which skills to install, and you approve, steer and judge what comes back.
The model sits in the middle, but the question starts with you, and so does the decision about whether the answer is good enough.
It never moved.
Across eight rings, not one weight changed. The invented schedule at the top and your real day at the bottom came from the same model. Everything that improved was built around it, and most of it is ordinary software written by people.
Takeaways
- The model is fixed at use. Everything that changes an answer on the day happens in the rings around it.
- Tools give reach; MCP makes reach portable. One connector, built once, works in every AI that speaks the protocol.
- Skills are know-how that waits. Only a one-line description is seen until the task calls for the rest.
- Memory is a saved note. Continuity comes from notes the harness hands back, not from the model remembering.
