Most people who use Claude every single day have no idea what’s happening under the hood. And honestly, that’s not their fault. Nobody sits down and explains the engine. You get the surface pitch: “it’s an AI assistant, it uses transformers, it can code.” The mechanics that explain why it behaves the way it does? Those stay hidden.
So when I stumbled on a post where a savvy professional mapped out 30 facts about how Claude really works, I stopped scrolling. The author put this together because they kept fielding the same questions from founders in their network. People running powerful tools they only half understand.
I love that framing, because it nails the real problem. When you don’t understand the engine, you drive the car wrong.
Why this gap matters
The original poster pointed out the exact frustrations most of us have felt. You get confused when Claude forgets things mid-conversation. You wonder why longer prompts sometimes make answers worse. You can’t figure out why it refuses one request but happily does another.
None of that is random. Every one of those quirks traces back to a design choice. Once you see the choices, the weird behavior starts making sense.
The core mechanics
Here’s how the expert broke down the foundation. Claude is a large language model, same category as ChatGPT and Gemini. The transformer at its heart uses attention to weigh how tokens relate to each other, all at once, not one by one.
That’s a key detail. It reads the whole input in parallel, not left to right like a human scanning a sentence. After the attention step, it outputs a probability distribution and picks the most fitting next token. That loop repeats, token by token, until the answer is finished.
Where Anthropic does things differently
This is the part that surprised me most. The creator laid out how Anthropic’s approach breaks from the usual playbook:
- Constitutional AI (CAI): training is guided by a written set of principles, not just human preference scores.
- Self-critique in the supervised phase: Claude reviews its own responses against that constitution.
- AI-scored RL phase: an AI model grades constitutional compliance to train a preference model.
- Baked-in safety: guardrails live inside training itself, not as bolt-on filters after generation.
- Authority hierarchy: operator system prompts outrank user messages, within Anthropic’s hard limits.
- Context housekeeping: internal reasoning tokens get stripped after each turn to protect the usable context budget.
- Extended thinking: a hidden scratchpad lets Claude reason step by step before it answers.
- No persistent memory: every conversation starts fresh, nothing carries over.
- Context rot: a bigger context helps with complex tasks but can drift on accuracy over time.
- Production tricks: quantization, KV-caching, and dynamic batching cut latency and cost.
The one takeaway that stuck
Model quality comes from architecture, data, and alignment working together. Not parameter count alone.
That single idea reframes a lot. The person who shared it made the point that this changes how you think about prompting, about tool use, and about what you should realistically expect from the model. Bigger isn’t automatically better. The recipe matters more than the raw size.
My strategic take
If you’re building with AI right now, this is the layer most people skip, and skipping it costs you. So here’s the move I’d recommend based on the author’s breakdown. Treat these facts as a checklist against your own workflow. Keep prompts tight because attention degrades on bloated context. Front-load the important instructions since context rot hits long threads. And stop expecting memory across chats, feed the model what it needs each time.
Understanding the engine genuinely changes how you build with it. The mechanics aren’t trivia, they’re the operating manual.
Go read the full post from the original creator for all 30 facts and the detail behind each one. Then send it to whoever on your team is running Claude without really understanding it. I’m curious which fact would surprise them most.