Most people who open Claude every single day couldn’t tell you what’s happening behind the screen. And that’s not their fault. Nobody sits them down and explains the engine, so they just keep driving the car and hoping for the best.
I came across a brilliant breakdown from an AI professional who mapped out 30 facts about how Claude really works. The original poster said they built the list because founders in their network kept asking the same confused questions over and over. I was genuinely hooked reading it, because it explains so many of the little frustrations we all hit.
Why understanding the engine matters
The author makes a point that stuck with me: when you don’t understand the engine, you drive the car wrong.
Think about the everyday head-scratchers:
- Claude seems to forget things in the middle of a conversation
- Longer prompts sometimes make the quality worse, not better
- It refuses some requests but happily does others
None of that is random. Once you see the mechanics, every one of those quirks starts to make sense.
The basic engine, explained simply
The creator lays out the foundation first. Claude is a large language model, the same family as ChatGPT and Gemini.
At its core sits a transformer that uses attention to weigh how tokens relate to each other, all at once. It reads the entire input in parallel, not left to right the way you and I read a sentence. Then it produces a probability distribution and picks the most fitting next token. That loop repeats, one token at a time, until the answer is done.
Simple enough. But this is where the expert points out that Anthropic takes a different road.
What Anthropic does differently
Here’s the part I found most interesting from the post. These are the design choices that shape how Claude actually behaves:
- Constitutional AI: training is guided by a written set of principles, not just human preference scores
- Self-critique: in the supervised phase, Claude checks its own responses against that constitution
- AI-scored alignment: in the reinforcement phase, an AI model rates constitutional compliance to train a preference model
- Baked-in safety: guardrails live inside training itself, not as filters slapped on after generation
- Authority levels: operator system prompts outrank user messages, within Anthropic’s hard limits
- Context hygiene: internal reasoning tokens get stripped after each turn to save usable context budget
- Extended thinking: a hidden scratchpad lets Claude reason step by step before it answers
- No memory: Claude has no persistent memory, so every conversation starts fresh
- Context rot: bigger context helps with complex tasks but can quietly degrade accuracy over time
- Production tricks: quantization, KV-caching, and dynamic batching cut latency and cost
Read that list again and the earlier frustrations click into place. Claude forgets because it has no memory. Quality drifts on long prompts because of context rot. The refusals trace back to constitutional training, not a mood.
The insight that changes how you build
Model quality comes from architecture, data, and alignment working together. Not parameter count alone.
That’s the line the original poster kept coming back to, and I think it’s the real takeaway. We’ve been trained to obsess over how big a model is. The expert reframes it: the magic is in how the pieces fit, not just the size of one of them.
That shift matters because it changes how you prompt, how you lean on tool use, and what you can fairly expect from the model in the first place.
My strategic recommendation
If you’re building anything with AI right now, treat this like operator training, not trivia. Spend 20 minutes learning the engine before you write your next 20 prompts. Match your habits to the mechanics: keep prompts tight because of context rot, don’t expect memory across chats, and lean on extended thinking for hard problems.
The people who understand the layer under the hood will simply get more out of the same tool than everyone still guessing.
The full breakdown from the creator has all 30 facts with more detail, so check out the original LinkedIn post and see which one surprises you most.