| name | swarm-orchestrator |
| description | Master multi-agent orchestration for RunSmart using Claude Code Agent Teams. Coordinates parallel specialists across development, marketing, QA, and operations. Use when tasks benefit from divide-and-conquer: feature builds, marketing campaigns, bug investigations, code reviews, or any work spanning multiple domains.
|
| metadata | {"short-description":"Orchestrate agent teams for parallel RunSmart development, marketing, and ops."} |
Purpose
Provides patterns and instructions for running Claude Code Agent Teams (swarms)
optimized for the RunSmart running coach PWA. Leverages Opus 4.6 native agent
teams with shared task lists, inter-agent messaging, and coordinated workflows.
When Claude should use this skill
- User requests parallel work across multiple files or domains
- Complex features spanning frontend, backend, AI, and tests
- Marketing campaign launches requiring copy, SEO, social, and analytics simultaneously
- Bug investigations benefiting from competing hypotheses
- Code reviews needing security, performance, and UX perspectives
- Any task where coordination overhead is justified by parallel speedup
Prerequisites
CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in settings (already configured)
- Opus 4.6 model for team lead (maximizes planning quality)
- Sonnet for teammates when speed > depth (cost optimization)
Swarm Patterns for RunSmart
Pattern 1: Feature Build Swarm
Best for new features (epics, stories) that span multiple layers.
Create an agent team for [feature name]:
- Architect teammate: design data schema, API routes, component structure
- Frontend teammate: implement React components and screens
- Backend teammate: implement API routes and database operations
- QA teammate: write tests and verify integration
Require plan approval from architect before frontend/backend proceed.
Task dependencies: Architect → (Frontend | Backend) → QA
Pattern 2: Marketing Launch Swarm
Best for coordinated marketing pushes across channels.
Create an agent team for RunSmart marketing launch:
- Copywriter teammate: homepage copy, app store description, email sequences
- SEO teammate: schema markup, meta tags, programmatic SEO pages
- Social teammate: Twitter/X threads, LinkedIn posts, Reddit content
- Analytics teammate: tracking setup, conversion funnels, A/B test config
Each teammate works independently, lead synthesizes into launch checklist.
Pattern 3: Bug Investigation Swarm
Best when root cause is unclear and multiple hypotheses exist.
Create an agent team to investigate [bug description]:
- Hypothesis A teammate: investigate [theory 1]
- Hypothesis B teammate: investigate [theory 2]
- Hypothesis C teammate: investigate [theory 3]
Have them challenge each other's theories. Update findings doc with consensus.
Pattern 4: Code Review Swarm
Best for thorough PR review across multiple dimensions.
Create an agent team to review the changes in [branch/PR]:
- Security reviewer: auth, input validation, XSS, injection
- Performance reviewer: bundle size, render cycles, database queries
- UX reviewer: mobile-first, PWA patterns, offline support
- Domain reviewer: running coach logic, safety guardrails, AI integration
Pattern 5: Release Preparation Swarm
Best for coordinating a production release.
Create an agent team for release preparation:
- Build teammate: run full build, fix errors, verify bundle
- Test teammate: run all test suites, fix failures
- Deploy teammate: prepare Vercel config, environment variables
- Docs teammate: update changelog, API docs, user-facing help
Effort Budgeting
- Team lead: Always Opus 4.6, maximum effort (strategic planning)
- Architecture/research teammates: Opus 4.6 or Sonnet (depends on complexity)
- Implementation teammates: Sonnet (speed-optimized execution)
- Review/QA teammates: Opus 4.6 (thoroughness matters)
File Conflict Prevention
- Each teammate should own distinct file sets
- Frontend teammate:
components/, app/ pages
- Backend teammate:
app/api/, lib/ utilities
- Test teammate:
*.test.tsx, *.test.ts files
- Never assign same file to multiple teammates
Coordination Rules
- Lead creates shared task list before spawning teammates
- Tasks have explicit dependencies (blocked-by relationships)
- No teammate verifies its own work — separate verification agent required
- Lead synthesizes results after all teammates complete
- Clean up team resources when done
Integration with RunSmart Skills
All teammates inherit project skills (running-coach-index, plan-generator, etc.)
via CLAUDE.md and .claude/skills/. Domain-specific teammates should reference
relevant skills:
- Feature teammates → running-coach-index contracts
- Marketing teammates → marketing/* skills
- Product teammates → product/* skills
- Solo decisions → solo-founder/* skills
Safety
- Never run agent teams for trivial single-file changes
- Monitor token usage — teams cost 3-5x a single session
- Always verify teammate outputs before merging
- Use plan approval mode for high-risk changes (database schema, auth)