Master Claude Code: Unlock AI‑Powered Terminal Coding

1. Agent Loop

When you give Claude a task, it goes through three intertwined stages: collect context, take action, and verify results. The loop uses tools such as file search, editing, and test execution, and adapts based on what it learns.

Model

Claude Code uses the Claude model (e.g., Sonnet for most coding tasks, Opus for complex architectural decisions). You can switch models with /model or claude --model.

Tools

Built‑in tools are grouped into four categories:

File operations : read, edit, create, rename, reorganize files.

Search : pattern matching, regex search, repository browsing.

Execution : run shell commands, start servers, run tests, use git.

Web : search the internet, fetch documentation, locate error messages.

Additional capabilities include sub‑agents, MCP extensions, hooks, and skill plugins.

2. What Claude Can Access

Claude Code runs in the terminal and can see:

Project : all files in the current directory and sub‑directories.

Terminal : any command you run (build tools, git, package managers, scripts, etc.).

Git status : current branch, uncommitted changes, recent commit history.

CLAUDE.md : a markdown file for persistent project‑specific context.

Extensions : MCP servers, skill plugins, sub‑agents, and Chrome‑based Claude integration.

Because Claude sees the whole project, it can work across files and branches, unlike a simple inline code assistant.

3. Using Sessions

Each Claude Code session is tied to a directory and stores the full conversation, tool usage, and results locally. Sessions are temporary; they do not persist across restarts unless you save context in CLAUDE.md. You can resume a session with claude --continue or claude --resume, fork a session with --fork-session, or use git worktree for parallel branches.

3.1 Cross‑branch work

Switching branches changes the visible files, but the conversation history remains, allowing you to continue the same discussion on a new code base.

3.2 Restoring or branching sessions

Resuming re‑attaches to the same session ID; forking creates a new ID while preserving the original history. Multiple terminals can share a session, but messages interleave, so for parallel work use --fork-session to give each terminal its own clean session.

claude --continue --fork-session

4. Checkpoints and Permissions

Claude Code creates a snapshot of each file before editing, allowing you to roll back with Esc twice. Checkpoints are local and only affect file changes; they do not cover external actions such as API calls or deployments, which is why Claude asks for confirmation before executing side‑effect commands.

Permissions control which operations Claude may perform without prompting. You can toggle modes with Shift+Tab:

Default : Claude asks before file edits and shell commands.

Auto‑accept edits : edits are applied automatically, but commands still require confirmation.

Plan mode : Claude uses read‑only tools to create a plan that you approve before any write actions.

Specific commands can be whitelisted in .claude/settings.json, preventing repeated prompts for trusted actions like npm test or git status.

5. Tips for Efficient Use

Ask Claude for help with concrete questions (e.g., “How do I set up a hook?” or “What’s the best way to structure CLAUDE.md?”).

Iterate: start with a rough prompt, then refine based on Claude’s output.

Interrupt Claude anytime with a correction and Enter; it will stop the current action and adjust its approach.

Provide precise initial prompts: reference specific files, constraints, and example patterns.

Supply verifiable expectations (test cases, UI screenshots, expected output) so Claude can validate its work.

Use Plan mode (press Shift+Tab twice) to analyze the codebase before making changes.

Delegate complex tasks to Claude as a sub‑agent rather than issuing low‑level commands; sub‑agents work with isolated context and return concise summaries.

These techniques help you get more accurate results from Claude Code and keep the conversation focused and efficient.