claude-code

name	claude-code
description	How to get authoritative information about Claude Code itself (the CLI, headless/print mode, SDK, settings, hooks, slash commands, MCP) and how to run it non-interactively from cron. Two sources of truth - the internal claude-code-guide agent and the online docs at code.claude.com/docs - plus a verified headless quick-reference. Use when a task needs Claude Code feature details, when automating Claude Code from a script/cron, or before answering any "can Claude Code ..." question (do not answer those from memory).
claude	please note - this skill exists in a public repo and should only be updated with specific written signoff by the updater. Keep skills/README.md (the skills index) in sync when any skill is added, renamed, removed, or its description changes.

claude-code

Reference for getting correct Claude Code information and for running Claude Code headless (e.g. from cron). Do NOT answer Claude Code feature questions from memory - the CLI changes fast. Use the two sources below, and verify flags against the local binary.

1. Authoritative sources (use these, not memory)

• Internal claude-code-guide agent - the primary source. It can read the local install and the live docs. Invoke via the Agent tool:
  • Agent(subagent_type="claude-code-guide", prompt="<specific question>")
  • Best for: feature questions, flag behavior, SDK usage, "how do I ...".
• Online docs - canonical host is code.claude.com/docs/en/ (docs.claude.com/... 301-redirects here):
  • Index of all pages (machine-readable): https://code.claude.com/docs/llms.txt
  • Overview: https://code.claude.com/docs/en/overview
  • Headless / programmatic (-p): https://code.claude.com/docs/en/headless.md
  • CLI reference (all flags): https://code.claude.com/docs/en/cli-reference
  • Settings / config: https://code.claude.com/docs/en/settings
  • Permission modes: https://code.claude.com/docs/en/permission-modes
  • Hooks: https://code.claude.com/docs/en/hooks
  • Skills: https://code.claude.com/docs/en/skills
  • Agent SDK (Python/TS): https://platform.claude.com/docs/en/agent-sdk/overview
  • Fetch a .md page with WebFetch when you need current detail.

2. Always verify flags locally

Flags and choices drift by version. Confirm before relying on them:

claude --version          # this box: see local install at ~/.local/bin/claude
claude --help | grep -- --<flag>

Verified on the local install (claude 2.1.197): -p/--print, --bare, --output-format (text default | json | stream-json), --allowedTools/--allowed-tools, --permission-mode (acceptEdits|auto|bypassPermissions|default), --model, --append-system-prompt[-file], --settings, --dangerously-skip-permissions, --continue, --resume. There is no --max-turns in this version.

3. Headless / cron quick-reference

claude -p "<prompt>"                         # non-interactive, prints result and exits
claude --bare -p "<prompt>" --allowedTools "Read,Bash"   # CI/cron: skip auto-discovery, pre-approve tools
claude -p "<prompt>" --output-format json | jq -r '.result'   # extract just the text result

• --bare skips hooks, skills, plugins, MCP, auto-memory, and CLAUDE.md - predictable for scripts. Recommended for cron (will become the -p default).
• --allowedTools "Read,Bash" pre-approves tools so the run never blocks on a prompt. Uses permission-rule syntax, e.g. Bash(git diff *).
• --output-format json payload includes total_cost_usd and a per-model cost breakdown; the text answer is in .result.
• User skills work in -p: put /skill-name in the prompt string (NOT available with --bare, which skips skill discovery). Interactive commands (/login, /config dialogs) do not work in -p.

Auth in headless (no interactive login)

• ANTHROPIC_API_KEY env var - required when using --bare (bare mode skips OAuth and keychain reads). API key from https://platform.claude.com.
• CLAUDE_CODE_OAUTH_TOKEN - long-lived token from running claude setup-token once interactively; works WITHOUT --bare (cannot be combined with --bare).
• apiKeyHelper in --settings JSON - for rotating credentials.

Cron environment (cron has a minimal env)

# minute hour day month weekday
0 4 * * 1,3,5  /home/youruser/path/to/task.sh >> /tmp/task.log 2>&1

#!/bin/bash
set -euo pipefail
export PATH="/home/youruser/.local/bin:/usr/local/bin:/usr/bin:/bin"
export HOME="/home/youruser"
export ANTHROPIC_API_KEY="..."        # or CLAUDE_CODE_OAUTH_TOKEN without --bare
out=$(claude --bare -p "<prompt>" --allowedTools "Read,Bash" --output-format json)
summary=$(printf '%s' "$out" | jq -r '.result')
# deliver, e.g. via the /email skill:
python3 ~/.claude/skills/email/send_email.py --to [e-mail] --subject "..." --body "$summary"

• Set PATH (cron strips it; the binary is at ~/.local/bin/claude), HOME, and the auth env var explicitly.
• Background Bash tasks Claude starts are killed ~5s after the result; background subagents/workflows are waited on (capped at 10 min, tune with CLAUDE_CODE_PRINT_BG_WAIT_CEILING_MS).
• Piped stdin is capped at 10 MB; for larger input, write a file and reference its path.

Note

Most cron tasks need NO AI call - deterministic scripts (curl/jq/python) are cheaper and cannot hallucinate. Reach for headless claude -p only when the task genuinely needs natural-language generation or judgement.