How Claude Code Actually Works
10 min readThe mental model. Context windows, how Claude reads your project, and why this is fundamentally different from using Claude in a browser.
"I really am mostly programming in English now · telling the LLM what code to write in words. It hurts the ego a bit. But the power to operate over software in large code actions is just too net useful."
Andrej Karpathy, co-founder of OpenAI
Claude Code is not a chatbot
When you use Claude in a browser, you're having a conversation. It can remember things across sessions with memory enabled, but it has no access to your files, can't run commands, and can't take actions on your behalf.
Claude Code is different. When you run claude from inside a project folder, it reads everything in that folder · your files, your notes, your code, your documentation. It can edit files, run commands, check for errors, and act on your behalf. It's an agent, not a chat interface.
The distinction isn't memory · it's agency. Claude Code actually does things in your environment, not just respond to you.
AI-Enhanced vs AI-Native
This is the gap most people never cross. Most use AI like a faster Google: type a question, get an answer, close the tab. Same workflow, slightly faster. The operators who pull ahead run their entire business through systems that already know their context and execute on their behalf.
| AI-Enhanced | AI-Native |
|---|---|
| Uses AI for one-off tasks | AI handles entire workflows |
| Re-explains context every session | AI has permanent business context |
| Waits for AI to respond | AI works while they sleep |
| Productivity boost: 2–3x | Productivity boost: 10–50x |
Claude Code is what makes the jump possible. It's not a chatbot · it's an agent with a harness. It reads and writes files, runs commands, browses the web, and connects to your tools. It operates on your actual business, not just in a chat window.
Context is everything. The more relevant context Claude has, the better its output. A session with a well-structured project folder, a good CLAUDE.md, and connected MCPs produces dramatically better results than an empty terminal. Every skill you install from the SkillTree map runs better when the context around it is rich.
The context window
Claude has a limited context window · the amount of information it can hold in mind at once. Think of it like short-term working memory. The longer a session runs, the more the context fills with previous exchanges.
When the context gets too full, Claude starts to lose track of earlier instructions and details. This is the "agent dumb zone." You'll notice it when Claude repeats itself, ignores earlier instructions, or makes mistakes it wasn't making before.
How to manage context:
- Run
/compactwhen your session gets long · it summarizes the conversation and frees up space. Do this proactively at around 50% context usage, not when things are already going wrong. - Run
/clearto start completely fresh when switching to a different task. - Keep each session focused on one task. Don't jump between unrelated problems in the same session.
How Claude reads your project
When you run claude from inside a folder, Claude can see and read every file in that folder. This is why project structure matters · a well-organized folder with clear file names gives Claude a far better map of your project than a messy one.
A recommended structure for any project:
/my-project
CLAUDE.md ← your instruction file (covered next)
/src ← code or scripts
/docs ← documentation, notes, SOPs
/assets ← any supporting filesYou don't need to follow this exactly · the principle is that folder names should describe what's inside them. Claude reads the names and structure to orient itself, not just the file contents.
This is also the foundation your brain sits on. Once you build your second brain, this same project-reading behaviour is what lets every skill pull your company context automatically.
Power-user moves worth knowing early
You don't need these on day one, but they change how the tool feels once you're running real work through it.
Three flags and commands that level you up
YOLO mode · let it run without permission popups. When you have a focused, reversible task, you can let Claude run fully autonomously:
claude --dangerously-skip-permissions "your task here"Always commit first so you can roll back: git add -A && git commit -m "checkpoint". If it goes sideways, git reset --hard HEAD. Never run it on production or unattended on anything irreversible. Give it an outcome ("fix all lint errors in src/"), not a vague wish ("make it better").
/btw · ask a side question mid-task. Type /btw followed by your question in the same message. It answers from what's already in context without interrupting the main task, without cluttering your history, and at near-zero token cost. It's read-only · for anything that needs to open files or search, use a subagent instead.
/rc · hand a running session to your phone. Start a task in VS Code, type /remote-control (or /rc), scan the QR code, and you can watch Claude work and send prompts from your phone while your machine keeps running. Nothing moves to the cloud · your phone is just a window into your local environment. Keep VS Code open; the session ends if you close it.