@rules/routing.md

Claude Code Skill

<output_language>

Default all user-facing deliverables, saved artifacts, reports, plans, generated docs, summaries, handoff notes, commit/message drafts, and validation notes to Korean, even when this canonical skill file is written in English.

Preserve source code identifiers, CLI commands, file paths, schema keys, JSON/YAML field names, API names, package names, proper nouns, and quoted source excerpts in their required or original language.

Use a different language only when the user explicitly requests it, an existing target artifact must stay in another language for consistency, or a machine-readable contract requires exact English tokens. If a localized template or reference exists (for example *.ko.md or *.ko.json), prefer it for user-facing artifacts.

</output_language>

Source of truth: https://code.claude.com/docs/en/cli-reference.

Defaults

Parameter	Default
Model selection	Use Claude Code CLI default unless the user explicitly asks for --model
Effort	Use CLI default unless the user explicitly asks for --effort
Permission mode	--permission-mode default
Headless mode	-p / --print
CI / scripted run	Add --bare so the call ignores local hooks, plugins, MCP, and CLAUDE.md
Resume target	claude --continue (-c) for the latest session in the current directory; claude --resume (-r) for a session by ID or display name

Do NOT ask the user for model or effort unless explicitly requested.

Routing

Use this skill when the request actually needs the claude CLI or a separate Claude Code session.

• Read rules/routing.md before building a command when the request might be out of scope.
• Route away to direct editing or another skill when the user wants generic writing, documentation cleanup, or local edits without needing the claude CLI itself.

Examples

Positive triggers:

• "Use Claude Code to review this repository and summarize the risks."
• "Run claude in print mode and analyze this architecture."
• "Continue the last Claude Code session and ask it to finish the patch."
• "Resume the auth-refactor claude session and apply the next fix."
• "Use claude --bare -p so this CI step does not pick up local hooks."

Negative triggers:

• "Rewrite this runbook for readability."
• "Create a new skill for our repo."

Boundary trigger:

• "Research Claude Code permissions and tell me what they do." Use this skill only if the user wants the claude CLI involved; otherwise route to research or direct documentation work.

Critical: Print Mode

Always use -p / --print for non-interactive Claude Code runs. Positional prompts without -p start the interactive REPL instead, so a script that omits -p will hang waiting for a TTY.

# Non-interactive (headless / SDK)
claude --permission-mode default -p "your prompt here"

# Interactive REPL (initial prompt only — does NOT exit)
claude "your prompt here"

-p is the canonical SDK/CI entrypoint. The CLI was previously called "headless mode"; the -p flag is unchanged.

Bare Mode for CI and Scripts

Add --bare for any scripted or CI invocation. It skips auto-discovery of hooks, skills, plugins, MCP servers, auto memory, and CLAUDE.md so the call returns the same result on every machine. Bare mode also skips the OAuth keychain, so it requires ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN, or an apiKeyHelper passed through --settings.

ANTHROPIC_API_KEY=$KEY claude --bare -p "Summarize the diff" --allowedTools "Read"

For pipelines that need a long-lived OAuth token instead of an API key, generate one with claude setup-token and export it as CLAUDE_CODE_OAUTH_TOKEN (note: bare mode does NOT read CLAUDE_CODE_OAUTH_TOKEN — use ANTHROPIC_API_KEY or apiKeyHelper when --bare is set).

Running a Task

Read references/recipes.md for concrete command recipes before changing permission modes, resuming a session, restricting tools, or adding extra directories.

Permission Mode Selection

Six modes are supported. Pick the loosest mode the task safely allows.

Flag	When to use
--permission-mode default	General Claude Code usage with normal approval prompts
--permission-mode plan	Read-only analysis or planning with no file changes or shell execution
--permission-mode acceptEdits	The user explicitly wants Claude Code to write files (also auto-approves common filesystem commands like mkdir, mv, cp)
--permission-mode auto	Long autonomous tasks where prompt fatigue matters; a server-side classifier blocks risky actions but does not replace review
--permission-mode dontAsk	Locked-down CI: only permissions.allow rules and the read-only command set may run; everything else is auto-denied
--permission-mode bypassPermissions	Containers / VMs only — equivalent to --dangerously-skip-permissions, requires explicit user approval, and offers no protection against prompt injection

Auto mode requires Claude Code v2.1.83+ and is gated by plan, admin policy, model, and provider; if the CLI reports it as unavailable, do not retry.

Command Discipline

• Start from claude --permission-mode default -p "your prompt here".
• Add --model <model> or --effort <level> only when the user explicitly asks.
• Use --output-format only when the user wants something other than text. Supported values: text (default), json, stream-json.
• For schema-validated structured output, pair --output-format json with --json-schema '<JSON Schema>'.
• Add --max-turns <N> and/or --max-budget-usd <dollars> for autonomous print-mode runs that must not loop forever.
• Use --fallback-model <name> in print mode when a graceful degrade beats a hard failure on overload.
• Use --add-dir <path> when the task needs files outside the launch directory. (--add-dir grants file access; it does NOT load .claude/ configuration from the added directory, except for .claude/skills/.)
• Restrict tools with --allowedTools, --disallowedTools, or --tools (use --tools "" to disable all built-ins, --tools "default" for all, or a comma list like "Bash,Edit,Read").
• Add to / replace the system prompt with --append-system-prompt, --append-system-prompt-file, --system-prompt, or --system-prompt-file.
• Load MCP servers with --mcp-config <file-or-json>; add --strict-mcp-config to ignore every other MCP source.
• Ask before using --dangerously-skip-permissions. Prefer --permission-mode acceptEdits for normal file edits.

Resuming a Session

# Latest session in the current directory
claude --continue -p "continue the previous task"   # short: claude -c -p "..."

# Specific session by ID or by display name
claude --resume "auth-refactor" -p "continue with this follow-up"   # short: -r
claude --resume <session-id> -p "continue with this follow-up"

# Pin a UUID so a script always reuses the same session
claude --session-id 550e8400-e29b-41d4-a716-446655440000 -p "..."   # must be a valid UUID

# Resume sessions linked to a pull request
claude --from-pr 123 -p "address review comments"

Use --continue for the latest conversation in the current directory. Use --resume for a specific session by ID or by display name (set with --name / -n, or by /rename mid-session). Use --session-id only with a real UUID — the CLI rejects other strings. When resuming, keep the existing session's behavior unless the user explicitly asks to change the model, effort, or permission mode. Add --fork-session only when the user wants to branch from the existing session instead of reusing it.

After Completion

• Summarize the result, including any warnings or partial output.
• Tell the user they can resume with claude --continue (-c), claude --resume <id-or-name> (-r), or claude --from-pr <pr>.
• Ask whether to continue, adjust the prompt, or switch back to direct work.

Critical Evaluation

Treat Claude Code as a colleague, not an authority.

• Trust your own grounded knowledge when you are confident.
• Verify disagreements with current docs or primary sources before accepting a claim.
• Remember that a separate Claude Code session can still be wrong or stale.
• Let the user decide when there is genuine ambiguity.

Authentication

The CLI reads credentials in this precedence order: cloud provider env vars (CLAUDE_CODE_USE_BEDROCK / _VERTEX / _FOUNDRY) → ANTHROPIC_AUTH_TOKEN → ANTHROPIC_API_KEY → apiKeyHelper → CLAUDE_CODE_OAUTH_TOKEN → subscription OAuth from /login.

• Browser login: run claude once and follow the prompt, or claude auth login (use --console for Console billing, --sso to force SSO).
• Inspect / sign out: claude auth status or /status from inside a session; claude auth logout or /logout to clear credentials.
• Long-lived OAuth for CI: claude setup-token prints a token; export it as CLAUDE_CODE_OAUTH_TOKEN (not read by --bare).
• Direct API: set ANTHROPIC_API_KEY for X-Api-Key, ANTHROPIC_AUTH_TOKEN for Authorization: Bearer through a gateway.

Error Handling

• command not found: claude: tell the user Claude Code CLI is not installed; they can run claude install stable after setup.
• Auth errors: confirm precedence above (e.g. an unset unset ANTHROPIC_API_KEY may be needed when a subscription is active), then re-run claude auth login or claude auth status to confirm.
• Permission blocks: retry with an appropriate --permission-mode (plan for read-only, acceptEdits for file edits) or adjust --allowedTools / --disallowedTools. Do not escalate to --dangerously-skip-permissions without explicit user approval.
• Session not found: run claude --resume without an argument to pick from a list, or switch to claude --continue for the current directory; --session-id requires a UUID.
• Auto-mode unavailable: this is gated by plan, admin policy, model, and provider — it is not a transient outage; fall back to default or acceptEdits.
• Invalid flag or model errors: check claude --help, then re-run with supported options. claude --help does not list every flag — consult the CLI reference for the full list.