Claude Code 101: Concept Map

How the ideas from the course fit together: one agent, one loop, one context window, and a set of ways to shape what goes into it.

Claude Code 101 at a Glance

Claude Code is an AI agent: software that interacts with its environment and acts to reach a goal, using an LLM that runs in a loop with tools.

How Everything Plugs into the Agentic Loop

The loop is the engine. Every other concept either feeds the context, provides tools, gates actions or runs work in a separate context. Hover or tap a box, or switch to the kitchen view to see the same ideas as a restaurant kitchen.

Your Prompt Interrupt · Steer · Add CONTEXT WINDOW 1 · Gather Context 2 · Take Action 3 · Verify Results Agentic Loop repeat until the result is verified, then stop every session CLAUDE.md project memory on demand Skills task-specific know-how summary only clean-context review Subagent own context window adds tools tool definitions MCP external tools & data gates actions Permissions manual · auto-accept · auto fires every time Hooks deterministic rules

From Prompt to Response: Under the Hood

Who talks to whom, in order, from the moment the user presses Enter to the response. This is a simplified view: solid arrows are calls, dashed arrows are replies, and boxes mark parts that repeat (loop), happen only sometimes (opt) or branch (alt). Hover or tap a step.

    What Lives in the Context Window

    The context window is Claude's working memory. Everything you type, every file it reads and every tool result adds to it. The proportions below are illustrative, not measured.

    Main Session

    System
    CLAUDE.md
    MCP tools
    Conversation + tool results
    free
    • CLAUDE.mdAlways loaded, every session
    • Skill descriptionsOnly name and description until a skill is used
    • MCP tool definitionsAlways loaded, even when unused. Above 10%, Claude switches to tool search
    • Conversation + tool resultsGrows with every prompt, file read and command
    • Subagent summaryThe only thing a subagent brings back
    • Free spaceWhen it runs out, auto-compaction kicks in
    exploration: reads, searches, tool calls
    Subagent's own window, discarded when it finishes
    → returns only →
    summary
    /context

    Inspect. Shows the size and what takes the most space. Nothing changes.

    /compact

    Summarize. Keeps a memory of the work so far and frees space. Use it to keep going on the same feature. Details may be lost.

    /clear

    Start fresh. Removes the conversation, so it can't bias a new feature. Anything Claude should remember belongs in CLAUDE.md.

    Explore → Plan → Code → Commit

    The daily workflow, with the concepts you use at each stage.

    Plan mode (read-only): the cheapest place to course-correct
    Your permission mode decides how much Claude asks
    Step 1

    Explore

    Give Claude the relevant context.

    • Plan mode
    • explore subagent
    • CLAUDE.md
    • web search
    Step 2

    Plan

    Agree on the steps and what "correct" looks like.

    • Plan mode
    • plan subagent
    • clarifying questions
    • success criteria
    Step 3

    Code

    Approve the plan, then iterate until it works.

    • permission mode
    • agentic loop
    • test suite
    • tools / MCP
    • hooks
    Step 4

    Commit

    Review the diff, get a second opinion, then ship.

    • /diff
    • /code-review (clean context)
    • /rewind
    • commit message
    Fix nowAsk whyLeave it

    Choosing the Right Customization

    What It IsWhen It LoadsContext CostGuaranteed?Where It LivesBest For
    CLAUDE.mdMarkdown onboarding notes added to your promptEvery session, automaticallyAlways present, so keep it compactNo, guidance
    • Project root (team, committed)
    • User config folder (just you)
    Project facts, conventions, repeated corrections
    SkillFolder with a SKILL.md, plus scripts and resourcesWhen its description matches your requestName and description only, until usedNo, guidance
    • .claude/skills in the repo (team)
    • ~/.claude/skills (personal)
    Task-specific know-how: review checklist, commit format, brand guide
    SubagentSeparate agent with its own system prompt and context windowWhen Claude delegates a taskOnly the summary it returnsNo
    • Markdown files with YAML frontmatter
    • Built-ins: general-purpose, explore, plan
    Exploration, research, clean-context review
    MCPOpen standard connecting external tools and dataTool definitions load up frontHigh: every tool definition, even unused onesNo
    • claude mcp add
    • Local / user / project (.mcp.json)
    Context outside the codebase: issue trackers, databases, apps
    HookCommand that runs on a lifecycle eventOn its event: PreToolUse, PostToolUse, Stop, …NoneYes, deterministic
    • settings.json
    • .claude/settings.json to share
    Must-happen rules: format, log, block rm -rf

    "I Want Claude To…"

    I want Claude to…
    always know something about this project
    →CLAUDE.md
    know how to do something, but only for certain tasks
    →Skill
    do something every single time, no exceptions
    →Hook
    reach a system outside the codebase
    →MCP (or a CLI)
    research or review without cluttering my context
    →Subagent