parallel-power

name	parallel-power
description	Use when the user asks "how do I run things in parallel", "should I use worktrees", or wants the playbook for multi-session and parallel execution — worktrees, fan-out, writer/reviewer splits, headless mode, queue-and-collect. Reference-style passive skill that surfaces examples inline.
when_to_use	Reach for this for explaining parallel patterns, before choosing between `/fan-out` vs `/worktree-team`, or when teaching. Do NOT use for dispatching the work — use `/dispatch` to pick the route or `/fan-out` and `/worktree-team` to execute it instead.
disable-model-invocation	true
allowed-tools	[]
logical	reference content surfaced explaining the right pattern; no execution side effects

Parallel Power: Do More at Once

Pattern 1: Parallel Worktrees

Work on multiple features simultaneously with zero conflicts:

claude --worktree feature-auth     # Isolated copy, own branch
claude --worktree bug-fix-123      # Another isolated copy
claude --worktree                  # Auto-generated name

Each worktree gets its own branch from origin/HEAD. Changes are isolated.

Switching worktrees: Use EnterWorktree with a path parameter to jump into an existing worktree — no need to create a new one each time.

Cleanup: Automatic when no changes were made. Worktrees whose PRs were squash-merged are also auto-removed.

Tip: Copy .env and other gitignored files by creating .worktreeinclude:

.env
.env.local
config/secrets.json

Pattern 2: Writer/Reviewer Split

Use SEPARATE sessions for writing and reviewing — fresh context prevents bias:

1. Session 1: Write the code
2. Session 2: claude --resume or new session → review the code
3. Fresh context means the reviewer isn't anchored to the implementation decisions

Pattern 3: Fan-Out Batch Processing

Process multiple files with the same operation:

# Generate task list
claude -p "list all controllers that need auth middleware" > tasks.txt

# Process each
while read file; do
  claude -p "add auth middleware to $file" --allowedTools "Read,Edit"
done < tasks.txt

Pattern 4: Orchestrator-Workers (Subagents)

For large features with independent subtasks:

• Create a plan with independent work items
• Ask Claude to dispatch subagents for each item
• Each subagent works in isolation, returns a summary
• Main session synthesizes results

Sweet spot: 3-5 parallel agents. More than that is hard to review.

That 3-5 ceiling is a human-review limit, not a system one. When the batch is large and mechanical — a codebase-wide sweep, a few-hundred-file migration, research cross-checked across many sources — a native dynamic workflow (trigger it with the ultracode keyword in your prompt) orchestrates many more subagents under the runtime, with adversarial verification built in, and runs in the background while your session stays free. Keep hand-rolled subagents when you want to review each result yourself; reach for a workflow when scale matters more than per-item review. Available on Claude Code builds that support dynamic workflows.

Tip: /loop runs a prompt or slash command on a recurring interval — pair it with a saved workflow for recurring orchestration.

Pattern 5: Session Per Concern

Don't mix unrelated tasks. Instead:

• claude -n "auth-refactor" — session for auth work
• claude -n "api-tests" — session for test writing
• claude -n "bug-4521" — session for a specific bug
• Resume any with claude --resume auth-refactor