Claude Code SOP

Claude Code SOP — Agent Orchestration Guide

Version: 2 (2026-02-21) | Claude Code: v2.1.50 | Auth: Anthropic subscription (not API key)

Overview

Claude Code is a full coding agent CLI with file editing, bash execution, subagents, skills, and multi-agent teams. It’s superior to sessions_spawn for serious builds because it iterates on errors, explores codebases, and coordinates multi-file changes.

Three Modes of Operation

Mode 1: Plan → Execute (RECOMMENDED for builds)

Phase 1 — Plan (read-only analysis):

cd /path/to/project
claude -p --permission-mode acceptEdits --model sonnet --max-budget-usd 1 \
  "Analyze the codebase. Create a detailed plan in PLAN.md.
   Be specific: file paths, component names, exact changes.
   ONLY create PLAN.md, do NOT modify existing files."

Phase 2 — Execute (full permissions, batched):

claude -p --dangerously-skip-permissions --model sonnet --max-budget-usd 3 \
  "Read PLAN.md. Execute items 1-5. Run 'npm run build' after. Fix errors. Commit.
   Do NOT ask questions. Just execute."

CRITICAL: Break large plans into batches of 5–7 items per run. 14+ items in one run risks hitting the 10-min timeout.

Mode 2: Stream-JSON Interactive (for monitoring + steering)

claude -p \
  --input-format stream-json \
  --output-format stream-json \
  --verbose \
  --dangerously-skip-permissions \
  --model sonnet

Receives NDJSON events on stdout (tool calls, thinking, results). Accepts NDJSON messages on stdin for follow-up instructions. Best used from tmux or a persistent shell.

Mode 3: Agent Teams (parallel workers)

CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 claude --teammate-mode tmux \
  --dangerously-skip-permissions --model sonnet \
  "Build X and Y in parallel. See PLAN.md."

Spawns team lead + parallel teammates in tmux panes. Max 5 agents. Requires: tmux installed, CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1.

Mandatory Rules

Project must be a git repo — git init if needed. Claude Code requires it.
Always include “Do NOT ask questions” — In -p mode, questions cause it to exit without doing work.
Always include build/verify step — “Run npm run build and fix any errors.”
Always include commit step — “Commit with message ‘X’.” Protects work from being lost.
Batch large tasks — Max 5–7 items per -p run. 10+ items risk timeout.
Set --max-budget-usd — Prevents runaway costs. $3–5 per batch.
Set timeout to 600s — Claude Code needs time for multi-file work.
Monitor progress — Check git diff --stat during runs to see what’s changing.
Update the user every 10 min — Never go silent on long-running tasks.

Available Skills

Skill	When to Use
`superpowers:brainstorming`	Before any creative/design work
`superpowers:writing-plans`	Creating implementation plans
`superpowers:executing-plans`	Executing a written plan
`superpowers:dispatching-parallel-agents`	2+ independent tasks
`superpowers:subagent-driven-development`	Multi-task execution
`superpowers:systematic-debugging`	Bug investigation
`superpowers:test-driven-development`	TDD workflow
`superpowers:verification-before-completion`	QA before marking done
`frontend-design:frontend-design`	UI/UX design work

Available Agents

Agent	Model	Purpose
`Explore`	haiku	Cheap codebase exploration
`Plan`	inherit	Read-only planning
`feature-dev:code-architect`	sonnet	Architecture design
`superpowers:code-reviewer`	inherit	Code review

Key Flags Reference

Flag	Purpose
`-p` / `--print`	Non-interactive, exit after response
`--permission-mode plan`	Read-only
`--permission-mode acceptEdits`	Auto-approve edits, prompt bash
`--dangerously-skip-permissions`	Full auto, no prompts
`--model sonnet` / `--model opus`	Model selection
`--max-budget-usd <n>`	Cost cap
`-c` / `--continue`	Resume last session in this dir
`--append-system-prompt`	Add custom instructions
`--agent <name>`	Use a named agent
`--teammate-mode tmux`	Multi-agent teams
`-w` / `--worktree [name]`	Git worktree for isolation

Claude Code vs sessions_spawn

Aspect	Claude Code	sessions_spawn
Interactive control	Full (can iterate)	One-shot only
Codebase awareness	Automatic	Must include in prompt
Error recovery	Iterates on failures	Must respawn
Multi-file coordination	Excellent	Hit or miss
Cost	Anthropic sub (rate-limited)	Provider API
Reliability	High (local process)	Announce can be missed

Use Claude Code for: Builds, refactors, multi-file features, anything that needs iteration. Use sessions_spawn for: Quick one-off tasks, research, non-coding work.

Standard Invocations

Agent teams (parallel/complex work):

CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 claude \
  --teammate-mode tmux \
  --dangerously-skip-permissions \
  --model <opus|sonnet per scope> \
  --max-budget-usd 5 \
  -p "Task brief. Use subagent-driven-development skill for parallel work.
      Use verification-before-completion skill before finishing.
      Use jCodeMunch MCP for code navigation (search_symbols, get_symbol, get_file_outline).
      Index with index_folder first if not already indexed.
      Do not read entire files. Commit. Do NOT ask questions."

Simple single-agent:

claude -p --dangerously-skip-permissions --model <model> --max-budget-usd 3 \
  "Task brief. Use verification-before-completion skill before finishing.
   Use jCodeMunch MCP for code navigation.
   Index with index_folder first. Do not read entire files. Commit. Do NOT ask questions."