Claude Code CLI introduction

Working loop

Choose a mode that matches the next action

1. Use Plan mode to inspect the project and receive an execution plan before source changes.
2. Approve edits only after the proposed scope and verification method are clear.
3. Use Bash mode for a command you have selected and want to run yourself.
4. Use Esc to interrupt a running loop and supply new information or a new boundary.

Check: the selected mode permits the required work and blocks actions outside the current task.

Permission modes · Interactive controls

Context management

Curate the working context for the current step

The context window contains instructions, selected files, tool outputs and conversation history available to the current model call.

1. Use /context to inspect what occupies the current window.
2. Name relevant files directly and exclude folders that cannot contribute evidence.
3. Use /compact when the same task continues after a long investigation.
4. Use /clear when a distinct task needs an empty working context. The saved session remains available for later resumption.

Check: every large context item contributes to the next decision or tool action.

Context window · How Claude Code works

Visual feedback

Turn screenshots into explicit acceptance checks

1. Capture the smallest stable region that demonstrates the issue.
2. Remove accounts, personal paths, tokens and confidential project data.
3. Describe the observed state and the required state in text.
4. Render the result at the same viewport and compare it with one acceptance criterion.

Check: the final screenshot demonstrates the requested state under the same conditions.

Common workflows

External systems

Add an MCP connection for a defined external need

The Model Context Protocol lets Claude Code use tools supplied by an external service or local server.

1. Name the external data or action required by the task.
2. Inspect the server source, requested credentials and available tools.
3. Separate read access from tools that create, edit or send information.
4. Add the server at the narrowest appropriate scope and inspect its state with /mcp.

Check: every exposed tool is required for the workflow, and its read or write effect is understood.

Official MCP guide

Reusable procedures

Use Skills for knowledge that should travel with a workflow

A Skill is a reusable instruction package for a defined class of tasks. Claude Code can load it when the task matches its description or when the user invokes it.

1. Stabilise the procedure and its completion check through a successful recurring workflow.
2. Define one concrete task class and the trigger that should load it.
3. Keep the main procedure in SKILL.md and supporting examples or scripts in separate resources.
4. Run the Skill on a fresh case and verify the artefact through its stated completion checks.

Check: the Skill removes repeated procedural prompting while keeping its evidence and checks inspectable.

Skills in the extension overview

Distribution

Package a stable extension when it must be reused

Plugins can bundle Skills, subagents, hooks and MCP servers for installation across projects or teams.

1. Develop a project-specific extension in .claude/ while its contract is still changing.
2. Package the components after their triggers, permissions and checks are stable.
3. Inspect a plugin and its included servers before installation.
4. Use a marketplace as a catalogue and install only the selected plugin.

Check: the plugin has a defined purpose, trusted source and reviewable component list.

Create plugins · Discover plugins

Delegated research

Delegate a clear research question to a fresh context

Subagents are useful when the direction is already known and one distinct question requires concentrated investigation or extensive file reading. Each subagent begins with a fresh, isolated context window; only its returned result enters the main conversation.

1. State one answerable question and why it matters to the main task.
2. Provide the required source frame, boundaries and return format.
3. Ask for findings, evidence and unresolved uncertainty.
4. Let the main agent compare results, integrate accepted findings and run the final checks.

Check: the returned summary answers the delegated question and cites the evidence required for integration.

Official subagent guide · Parallel agent patterns

Parallel work

Match the worker and workspace to the task

A subagent supports one main session. A parallel session is an independently supervised working thread. A Git worktree gives a writing session its own checkout so concurrent edits do not collide.

1. Use a subagent for a bounded investigation whose detailed exploration should remain outside the main context.
2. Use a separate session when an independent task needs direct supervision and its own context.
3. Start a writing session with claude --worktree <name> when another session edits the same repository.
4. Compare the outputs, resolve conflicting assumptions and run shared checks before integration.

Check: each worker owns an independent question or file area, and concurrent edits cannot overwrite one another.

Parallel agent patterns · Claude Code worktrees

Optional control surface

Continue a local session from a remote interface

Remote Control is a research-preview route to the same local Claude Code session. The browser or mobile app controls it; the working directory, tools and permissions remain on the host machine.

1. Check the current official availability, account and organisation requirements.
2. Start with claude remote-control, claude --remote-control or /remote-control inside a running session.
3. Open the displayed link or QR code on the intended remote device.
4. Confirm the session and working directory before approving or interrupting an action.

Check: both interfaces show the same conversation and project while execution remains on the local host.

Official Remote Control guide

Next routes

Continue with the mechanism your work requires

Check: select one route because it resolves a concrete need in the current project.