Skip to main content
Strip Claude down to the skeleton and there are exactly three parts. A brain that predicts. Hands that act. A memory that decides what the brain gets to work with. Every feature with a name (models, tools, connectors, projects, agents, MCP) is one of these three parts under a different label. Hold the anatomy and none of them will ever confuse you.

​
Take the machine apart

The fastest way to understand a machine is to break it. Pull a part and watch what happens to the output. Put it back. Pull two.
What the wreckage teaches:
  • Only the brain thinks, and it never acts. When something has to actually happen in the world (a search, a calculation, an edit to a real file), the brain can only ask the hands to do it.
  • Only the memory is yours. You do not train the model and you do not build the tools. What you control is what sits in the window: what you say, what you attach, what you leave lying around in the conversation.

​
The sentence to keep

One line carries this entire page: models predict, tools act, and context constrains.
Models predictEvery word is a forecast of the next. Brilliance and hallucination are the same move at different confidence levels.
Tools actFresh facts, exact numbers, real actions. Anything that must be true comes from a tool call, not from a training memory.
Context constrainsThe model continues from whatever it is given. Rich context narrows prediction toward your answer. Empty context leaves it at everyone’s.

​
Same request, four machines

Here is one request a real operations lead might send. The brain is identical in every quadrant of this map. The only thing that changes is what it was given to work with.
Watch the fluency gauge while you click. It never moves. The generic guess and the exact, checkable answer arrive with the same polish and the same confidence, which means “sounds right” tells you nothing. Grounded and exact are the gauges you actually manage, and you manage them before you hit send.

​
Blame a part, not the machine

“The AI was wrong” is not a diagnosis. When an output disappoints you, walk the anatomy in order:
1

Did it have the memory?

Would a sharp colleague have produced this with exactly what you gave them, meaning two sentences and no files? If yes, generic was the correct output. The fix is context: attach the source, state the goal, define what good looks like.
2

Did it have the hands?

Did the task need fresh facts, exact numbers, or your real systems? If it did and no tool touched it, the answer was always going to be predicted. The fix is equipment: search, code, a connector, and asking for the work to be shown.
3

Was it ever prediction work?

Some jobs need authority, accountability, or knowledge that exists only in someone’s head. No amount of context or tooling fixes those. That short list is covered in when not to use Claude.
Most failures die at the first question. Nearly all the rest die at the second. The model itself is the last suspect, not the first, which is the opposite of how most people read a bad answer.
Think of the last time an AI answer genuinely disappointed you. Diagnose it out loud in the anatomy’s terms. “The model is dumb” is not an available answer until the other two parts have been cleared first.

The AI fluency loop

The habit that runs this machine well: assign, equip, inspect, iterate, ship.