Common Claude Code Commands: Interactive Sessions, Recovery, Scripts, and Structured Output

Running claude directly opens a persistent session for exploration, editing, and multiple rounds of feedback. Adding a prompt to the launch command starts with a specific task; -p prints a result in non-interactive mode for scripts and CI.

Validate prompts and permissions interactively before automating them. No one can correct a non-interactive run in real time, so its task boundaries and allowed tools must be stricter.

Enter slash commands in the Claude Code interactive interface, not directly in the shell. /help lists commands for the installed version, /status checks the model and authentication, /context shows context sources, and /permissions manages permissions.

The command set changes between versions. Treat /help as the authority for the installed version rather than presenting a static command list as a permanent interface.

claude --continue resumes the most recent session in the current directory. claude --resume opens a session selector scoped to the current worktree by default, with an option to expand to other projects. Give complex tasks searchable names with /rename.

Restoring a session restores the conversation and related state, but the current working tree may have changed. Have Claude recheck the Git status and key files before continuing.

Non-interactive mode supports plain text, JSON, and streaming JSON. A single JSON object works when parsing the final result; stream-json supports real-time event consumption and requires verbose output.

Do not use regular expressions to infer fields from natural-language output. Select a structured format, then validate required fields and exit status downstream.

Expose only the tools a batch task needs. --tools controls which built-in tools are available; --allowedTools only lets matching tools run without approval and does not hide other tools. Pass an empty --tools value for pipeline tasks that need no tools.

--safe-mode is useful for isolated reviews that use an existing subscription or OAuth login. The stricter --bare does not read OAuth or the keychain and only works when ANTHROPIC_API_KEY or apiKeyHelper is explicitly provided.

Calibrate prompts, permissions, and output on two or three samples before increasing the batch size. Parallel processing increases cost and can create conflicts in the same worktree, so editing tasks are better isolated in separate worktrees.

claude doctor checks installation and settings from the shell; claude --debug-file <path> writes detailed startup, MCP, or gateway diagnostics to the specified file. With --debug alone, the default log is stored at ~/.claude/debug/<session-id>.txt.

When troubleshooting, only change one variable at a time and keep the complete error text. Model errors, authentication errors, and network errors have different repair paths.