batch-claude-loop

Batch Claude Loop

ℹ️ First, set expectations: before anything else, print one short line with this skill's version and its most recent change — the top entry of metadata.changelog above — e.g. batch-claude-loop vX.Y — recent: <summary>. Then proceed.

Purpose

Orchestrate multiple headless Claude Code invocations in a loop — each with clean context, different input, and structured output collection. This skill builds and executes the batch script; individual worker calls run independently via claude -p.

Process

Step 1: Clarify the Batch Job

Determine from the user:

1. What to iterate over — file list, array of prompts, lines from a file, API results, etc.
2. What each worker should do — the prompt template with a {{ITEM}} placeholder
3. Which tools workers need — for --allowedTools (e.g., Read,Bash,Edit)
4. Worker directory — which directory should workers run in (default: current directory)
5. Parallelism — serial (default) or parallel (with max concurrency)
6. Output format — text, json, or stream-json

Step 2: Configure Worker Flags

Select flags for each headless invocation:

# Required flags
claude -p "${PROMPT}"              # Headless mode

# Recommended flags
--output-format json               # Structured output (parse with jq)
--allowedTools "Read,Bash"         # Pre-approve tools (no permission prompts)
--max-turns 5                      # Prevent runaway execution

# Optional flags based on needs
--bare                             # Skip all auto-discovery (fastest, most isolated)
                                   # Omit if workers need skills/MCP/CLAUDE.md
--max-budget-usd 1.00              # Cost ceiling per worker
--no-session-persistence           # Don't save session history
--append-system-prompt "..."       # Inject extra instructions

Decision: --bare vs normal mode

• Use --bare when workers are self-contained (just a prompt + tools)
• Omit --bare when workers need project context (CLAUDE.md, skills, MCP servers)