Key: Claude has no permanent memory. It only "sees" what's inside the context window for that conversation. If you don't tell it, it doesn't know.
When you use Claude, there are three active zones at once: the extensions that expand its capabilities, your project with all your files, and the context window where Claude processes everything.
Skills, MCP, Plugins and Marketplace. Optional modules that expand what Claude can do.
The workspace. Your files, instructions and configuration (CLAUDE.md) live here.
Claude's "workbench". Only processes what fits there (~200,000 tokens). If it fills up, it forgets the oldest.
The result. Can be text (md), document (html/pdf/docx) or structured data (json).
System prompt — Defined by Anthropic. Sets Claude's base limits.
User preferences — Your personal CLAUDE.md with your global rules.
Project config — Specific instructions for this particular project.
Files and content — The documents you've uploaded or that are in the project.
Generated output — The responses Claude has already given in this session.
The CLAUDE.md file tells Claude how to behave. There are 4 levels. The lower in the hierarchy, the more specific and with higher priority in context.
📌 Easy rule: The lower on the ladder, the more authority in that specific context. The FOLDER level can override the USER for that specific folder.
Claude can respond in three main format types. Choose the one that best suits your need.
| Format | When to use it | Level |
|---|---|---|
| ↑ md — Markdown | Quick text with basic structure: bold, lists, headings. The most readable in chat. | Basic |
| ↑↑ html / pdf / docx | Visual deliverables with design: tables, images, layout. Ideal for reports or training. | Medium |
| ↑ json — Data | Structured data for another app or script to process automatically. | Advanced |
A Skill is a packaged process in a folder. You teach Claude to do a repetitive task once — and it executes it automatically whenever you need it.
💡 Analogy: A Skill is like an app on your phone. You install it once and run it when you need it. You define what it does and how.
name: — The skill name. Claude uses it to identify it.
description: — When and what it's used for. Claude reads it to know whether to activate it.
prompt: — The exact instructions Claude will follow when executing the process.
Creates new skills from scratch or improves existing ones.
Designs autonomous agents that execute complex tasks in multiple steps.
Packages MCPs + Skills into an installable and reusable plugin.
Claude's "workbench". Everything it can see in that conversation. Has a maximum size (~200k tokens).
Text file with instructions for Claude. Defines tone, rules, and protocol. Exists in 4 levels.
Model Context Protocol. The "universal plug" that connects Claude with APIs, databases and external services.
Process packaged in a folder. You define it once, Claude executes it whenever needed.
Installable bundle combining MCPs + Skills. Adds capabilities to Claude without manual configuration.
The workspace where your files, instructions and settings live. Claude reads it in full before responding.