Skill

swarm-skill

Decompose complex tasks into parallel subtasks and coordinate an Agent Team of teammates to complete them concurrently. Use when the user describes a multi-part development task or explicitly requests task splitting, parallelization, or uses keywords like team, swarm, parallel, 拆分, 并行, 分工. Not for single-step tasks (typo fix, single file edit, answering a question).

From swarm-plugin

Install

Run in your terminal

npx claudepluginhub istarwyh/agent-plugins --plugin swarm-plugin

Tool Access

This skill uses the workspace's default tool permissions.

Supporting Assets

View in Repository

references/patterns.md

scripts/check_env.sh

scripts/discover_skills.sh

Skill Content

Similar Skills

skill-lookup

Searches, retrieves, and installs Agent Skills from prompts.chat registry using MCP tools like search_skills and get_skill. Activates for finding skills, browsing catalogs, or extending Claude.

prompts.chat

157.6k

prompt-lookup

Searches prompts.chat for AI prompt templates by keyword or category, retrieves by ID with variable handling, and improves prompts via AI. Use for discovering or enhancing prompts.

prompts.chat

157.6k

Agent Development

6 files

Guides agent creation for Claude Code plugins with file templates, frontmatter specs (name, description, model), triggering examples, system prompts, and best practices.

plugin-dev

83.2k

Stats

Parent Repo Stars2

Parent Repo Forks1

Last CommitApr 6, 2026

Actions

View Source View Plugin View on GitHub View README

Swarm — Task Decomposition & Parallel Execution

Environment Check

Run scripts/check_env.sh first. Handle by exit code:

Exit 0: Ready. Proceed to skill discovery.
Exit 1: Version too low. Tell user to run claude update. Stop.
Exit 2: Agent Teams not enabled. Read ~/.claude/settings.json, add "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" to the "env" object (create if absent, preserve existing config), write back. Tell user to restart Claude Code. Stop.

Skill Discovery

Build the available skill inventory from two sources:

Session skills: Parse the <system-reminder> in current context for listed skills (name + description).
Filesystem commands: Run scripts/discover_skills.sh to scan ~/.claude/commands/ and .claude/commands/.

Merge into a flat list. Only use skills that actually appear — never assume.

Teammates auto-load the same skills as the lead. No extra config needed.

Workflow

Check env → Discover skills → Decompose → Create team + tasks → Spawn teammates → Monitor → Deliver

1. Decompose

Analyze the user's request. Split into subtasks that can run in parallel or pipeline.

Split when: parallelizable, different expertise, or pipelineable. Don't split: tightly coupled edits, single-file changes, heavy shared context.

See references/patterns.md for decomposition patterns and role-to-skill matching.

2. Create Team & Tasks

Create an Agent Team, then create Tasks. Each Task must be self-contained: clear goal, file paths, completion criteria, project conventions (from CLAUDE.md).

Use addBlockedBy for dependencies — blocked teammates wait automatically.

3. Spawn Teammates

Spawn independent teammates in parallel (single message, multiple Agent calls)
Dependent teammates also spawn immediately — they self-wait
All use run_in_background: true

Build prompts using the template in references/patterns.md. Only inject actually-discovered skills.

Agent type: default for code, "Explore" for research, "Plan" for design.

4. Monitor & Coordinate

Verify output on completion; SendMessage corrections if needed
Resolve conflicts if two teammates edit the same file
On failure: analyze cause, guide retry via SendMessage

5. Commit (optional)

Create a committer teammate addBlockedBy all dev/test tasks. Inject commit skill if discovered; otherwise use git directly. See references/patterns.md for prompt templates.

6. Deliver

Run tests to verify changes
Summarize each teammate's output to the user
Shut down teammates and clean up the team