Parallel Tasks

One flat team. All agents (captains and engineers) are teammates in the same team. Naming convention groups them by worktree:

Coordinator (you — delegate mode, thin dispatcher only)
  └── TeamCreate("parallel-work")
       ├── task1-captain                ← worktree 1 (autonomous)
       ├── task1-eng1                   ← spawned on captain's request
       ├── task1-eng2                   ← spawned on captain's request
       ├── task2-captain                ← worktree 2
       ├── task2-eng1
       └── task3-captain                ← worktree 3 (solo, no engineers)

Key constraints from the agent teams docs:

• No nested teams — teammates cannot spawn their own teams or teammates.
• Only the coordinator can spawn teammates — captains message the coordinator with staffing requests; the coordinator spawns them.
• One team per session — everything in a single flat team.
• All teammates can message each other — engineers in the same worktree coordinate directly via SendMessage (critical for file ownership).

Verification subagents (code-reviewer, code-simplifier) are spawned via the Task tool without team_name — they're one-shot workers that report results back to the captain.

Tools Used

• TeamCreate / TeamDelete — create and tear down the single team.
• Task — spawn teammates (with team_name) and verification subagents (without team_name).
• TaskCreate / TaskList / TaskUpdate — manage the shared task list.
• SendMessage — DMs between any teammates, shutdown requests, staffing requests.

Workflow

Phase 1: Setup (Coordinator)

1. Detect worktree base directory — check in priority order:

   # 1. Check for existing directories (.worktrees takes priority)
   ls -d .worktrees 2>/dev/null || ls -d worktrees 2>/dev/null
   
   # 2. If neither exists, check CLAUDE.md for a preference
   grep -i "worktree" CLAUDE.md 2>/dev/null
   
   # 3. If nothing found, ask the user where to put worktrees
   

   If creating a new project-local directory, verify it's in .gitignore:

   git check-ignore -q worktrees 2>/dev/null
   # If not ignored, add to .gitignore before proceeding
   

2. Detect the base branch and ensure it's up to date:

   BASE_BRANCH=$(git symbolic-ref refs/remotes/origin/HEAD 2>/dev/null | sed 's@^refs/remotes/origin/@@') || BASE_BRANCH="main"
   git checkout "$BASE_BRANCH" && git pull origin "$BASE_BRANCH"
   

3. Create worktrees — one per task, all branching from the base branch:

   git worktree add {worktree-base}/{task-name} -b feature/{task-name} "$BASE_BRANCH"
   

4. Create the team:

   TeamCreate(team_name="parallel-work", description="Parallel task execution across {N} worktrees")
   

5. Create a task for each worktree in the shared task list:

   TaskCreate(subject="Implement {task-name}", description="{full task description}", activeForm="Implementing {task-name}")
   

Phase 2: Dispatch Captains (Coordinator)

Switch to delegate mode (Shift+Tab) so you stay focused on coordination and don't start implementing tasks yourself.

Spawn one captain per worktree. Captains are autonomous — they plan, build, and deliver without needing coordinator approval.

Task tool settings for each captain:

• subagent_type: "general-purpose"
• mode: "bypassPermissions"
• team_name: "parallel-work"
• name: "{task-name}-captain"

Every captain prompt MUST include:

1. The working directory constraint (with the actual resolved path):

   ## CRITICAL: Working directory
   You MUST work ONLY in {worktree-base}/{task-name}/. This is an isolated git
   worktree already on branch feature/{task-name}.
   
   ALL file paths must be under {worktree-base}/{task-name}/.
   Do NOT touch the main working directory.
   The branch is already created. Do NOT run git checkout or git branch.
   

2. The task description from the user.

3. Any user-provided quality check overrides.

4. The coordinator's name for messaging back:

   ## Coordinator
   The coordinator's name is "{coordinator-name}". Message them via SendMessage to:
   - Request teammates (engineers, etc.) — include role names, sub-tasks, and file ownership
   - Report your PR when complete
   - Ask for help if blocked
   
   Your task has already been assigned to you in the shared task list — update its
   status to in_progress when you start building, and completed when your PR is created.
   

   To find your own name for this placeholder, read the team config at ~/.claude/teams/parallel-work/config.json after creating the team.

5. The full Captain Playbook (below) — paste it into each captain's prompt.

After spawning all captains, use TaskUpdate to assign each task to the corresponding captain.

Phase 3: Handle Staffing Requests (Coordinator)

Captains are autonomous — they orient, plan, and may start building immediately. If a captain needs teammates, they'll message you with a staffing request.

When you receive a staffing request, spawn the requested roles:

• subagent_type: "general-purpose"
• mode: "bypassPermissions"
• team_name: "parallel-work"
• name: "{task-name}-eng1", "{task-name}-eng2", etc.

Each teammate's prompt MUST include:

• The same worktree path constraint as the captain
• Their specific sub-task (from the captain's request)
• Which files they own (to avoid conflicts)
• Their captain's name for coordination: "Message {task-name}-captain if you have questions or need to coordinate file access."

Also create TaskCreate entries for each teammate's sub-task and assign them.

Phase 4: Monitor (Coordinator)

Captains implement and report back:

1. Task list — Use TaskList to see all tasks at a glance. Captains and engineers update status as they progress.
2. Messages — Captains send you messages when they complete (with PR URLs) or need help. Delivered automatically.
3. Shutdown — As each captain reports a PR, shut down the captain's team:

   SendMessage(type="shutdown_request", recipient="{task-name}-eng1")
   SendMessage(type="shutdown_request", recipient="{task-name}-eng2")
   SendMessage(type="shutdown_request", recipient="{task-name}-captain")
   

4. Track progress in a summary table:

   | Task | Captain | Engineers | Status | PR |
   |------|---------|-----------|--------|----|
   | item-4-hybrid-search | task4-captain | task4-eng1, task4-eng2 | PR created | #42 |
   | item-7-merge-suggestions | task7-captain | (solo) | In progress | — |
   

Phase 5: Cleanup (Coordinator)

Once all teammates have shut down (TeamDelete will fail if any are still active):

1. If a teammate rejects a shutdown request, wait for them to go idle and retry. If they persistently reject, message them explaining the work is complete.
2. TeamDelete to remove the team and its task list.
3. Remove worktrees:

   git worktree list
   git worktree remove {worktree-base}/{task-name}  # for each
   # If removal fails due to uncommitted changes:
   git worktree remove --force {worktree-base}/{task-name}
   

4. Present final summary with all PR URLs.

Captain Playbook

Paste this entire section into each captain's prompt. It defines how a captain operates inside their worktree.

Step 1: Orient

Run project setup — auto-detect from project files in your worktree:

# Node.js
[ -f package.json ] && npm install
# Python (uv)
[ -f pyproject.toml ] && uv sync --all-extras
# Python (pip)
[ -f requirements.txt ] && pip install -r requirements.txt
# Rust
[ -f Cargo.toml ] && cargo build
# Go
[ -f go.mod ] && go mod download

Read the project's CLAUDE.md (if it exists) in your worktree root. Extract:

• Quality commands — lint, typecheck, test commands (e.g., uv run ruff check, uv run pyright, uv run pytest)
• Coding standards — line length, import style, type annotation policy, suppression rules
• Project structure — where code lives, key patterns, naming conventions

These will be used in verification gates and in your plan.

Step 2: Plan

Analyze your assigned task:

• Read relevant existing code to understand the codebase
• Break the task into sub-tasks if it has independent parts
• Identify files that will be created or modified
• Note any dependencies between sub-tasks

Decide on staffing:

If you need teammates, message the coordinator with your staffing request. Include role names, sub-tasks, and file ownership for each. Then wait for the coordinator to confirm they've been spawned before proceeding.

SendMessage(type="message", recipient="{coordinator-name}",
  content="Staffing request for {task-name}:
    - {task-name}-eng1: {sub-task}, owns {files}
    - {task-name}-eng2: {sub-task}, owns {files}",
  summary="Staffing request for {task-name}")

If solo, proceed directly to Step 3.

Step 3: Build

If solo: Implement directly. Follow the project's coding standards from Step 1.

If you have engineers: Coordinate via SendMessage and TaskList:

• Message engineers with clarifications or priority changes
• Engineers can message you and each other directly
• Monitor TaskList for sub-task completion
• Resolve any file ownership conflicts immediately

Git workflow: Engineers edit files but do NOT commit. You (the captain) make the final commit in Step 5 after all verification gates pass. This avoids sequencing issues with multiple commits from different agents in the same worktree.

IMPORTANT: Wait for ALL engineer sub-tasks to be marked completed in TaskList before starting Step 4. Running verification on partially-complete work wastes time and produces misleading results.

Step 4: Verify (Four Gates)

All four gates must pass before creating a PR. Do not skip any gate.

Gate 1: Project Quality Checks

Run every quality command found in CLAUDE.md:

# These are EXAMPLES — use whatever the project's CLAUDE.md specifies
{lint command}      # e.g., uv run ruff check src/
{typecheck command}  # e.g., uv run pyright src/
{test command}       # e.g., uv run pytest

Fix any failures before proceeding. If the user provided additional quality check overrides, run those too.

Gate 2: Code Review

Spawn a code-reviewer subagent (Task tool, no team_name) on your changes:

Task(
  subagent_type="superpowers:code-reviewer",
  prompt="Review the changes on branch feature/{task-name} in {worktree-path}.
    What was implemented: {task description}
    Requirements: {original task description from user}
    BASE_SHA: {sha before your changes}
    HEAD_SHA: {current HEAD}
    Description: {brief summary of changes made}"
)

If the superpowers:code-reviewer subagent type is unavailable, review the diff yourself.

Act on feedback:

• Critical/Important issues — fix immediately, re-run Gate 1
• Minor issues — fix if straightforward, otherwise note in PR description

Gate 3: Code Simplification

Spawn a code-simplifier subagent (Task tool, no team_name) on modified files:

Task(
  subagent_type="code-simplifier:code-simplifier",
  prompt="Simplify and refine the recently modified code in {worktree-path}
    for clarity, consistency, and maintainability. Focus only on files
    changed on this branch. Preserve all functionality."
)

If the code-simplifier:code-simplifier subagent type is unavailable, review the code for simplification opportunities yourself.

If the simplifier makes changes, re-run Gate 1 to confirm nothing broke.

Gate 4: Requirements Verification

Re-read your original task description. For each requirement:

• Verify with concrete evidence (test output, file contents, command output)
• Do NOT claim completion based on memory or confidence — run the verification
• If any requirement is not met, go back and implement it

Only proceed when ALL gates pass.

Step 5: Deliver

cd {worktree-path}
git add -A
git commit -m "$(cat <<'EOF'
{commit message — summarize what was built and why}

Co-Authored-By: Claude <[email protected]>
EOF
)"
git push -u origin feature/{task-name}
gh pr create --title "{PR title}" --body "$(cat <<'EOF'
## Summary
{1-3 bullet points}

## Verification
- [ ] Quality checks passed (lint, typecheck, tests)
- [ ] Code review — {summary of findings and fixes}
- [ ] Code simplification — {summary of changes}
- [ ] Requirements verified against original task

## Test plan
{how to test the changes}

🤖 Generated with [Claude Code](https://claude.com/claude-code)
EOF
)"

Report the PR URL back to the coordinator (use the name from your prompt):

SendMessage(type="message", recipient="{coordinator-name}", content="PR created: {url}", summary="PR ready for {task-name}")

Then mark your task as completed via TaskUpdate and go idle. The coordinator will shut you down.

Rules

• Never let multiple agents edit the same files — captains must assign clear file ownership
• Always create worktrees BEFORE spawning any teammates
• Always run all four verification gates before creating a PR
• Only the coordinator spawns teammates — captains message the coordinator with staffing requests
• Coordinator is a thin dispatcher — spawn agents, handle staffing requests, monitor for PRs
• Captains are autonomous — they plan, build, and verify without needing coordinator approval
• Captains must read CLAUDE.md — do not hardcode project-specific commands in this skill
• Teammates may ignore shutdown requests while mid-turn — retry when they go idle
• Shut down engineers BEFORE their captain (captains may need to do final integration)
• Engineers in the same worktree MUST coordinate file access via SendMessage

Example Invocation

User: "Run these 3 tasks in parallel:

1. item-4-hybrid-search: Add BM25+vector hybrid search with RRF fusion scoring
2. item-7-merge-suggestions: Implement merge suggestion system with confidence thresholds
3. item-14-halfvec: Migrate embeddings from vector to halfvec for 50% storage reduction"

-> Detects worktrees/ directory, creates 3 worktrees, creates team, spawns 3 captains. -> Captains orient, plan, and request engineers as needed. Coordinator spawns requested engineers. -> Captains coordinate their teams, run 4 verification gates, create PRs. -> Coordinator collects 3 PR URLs, shuts down all teammates, cleans up, presents summary.