harness-for-yall

Multi-agent harness for Claude Code: 26 agents, 15 skills, 5 teams.

한국어 문서

What is this?

A pre-configured set of Claude Code agents organized into 5 specialized teams. Each team uses a different multi-agent orchestration pattern optimized for its purpose.

Teams

Plugin	Pattern	Agents	Skills	What it does
dev-pipeline	Pipeline	5	1	Full feature dev: planner → FE+BE parallel → reviewer gate → QA
review-pipeline	Fan-out / Fan-in	5	1	Code review: 3 parallel screeners → moderator → judge (SARIF output)
fe-experts	Expert Pool	5	5	Frontend: architect → implementer / styler → perf + tester
be-experts	Pipeline + Expert Pool	7	5	Backend: architect → impl + validator → resilience / provider / security → tester
explore-team	Hierarchical Delegation	4	3	Codebase exploration: scout(opus) → hypothesizer → evidence → synthesizer

Install

Option 1 — npx

# All plugins
npx harness-for-yall

# Pick specific teams
npx harness-for-yall fe-experts be-experts

# Preview first
npx harness-for-yall --dry-run

# Overwrite existing
npx harness-for-yall --force

Copies agents/skills to ~/.claude/. Zero dependencies.

Option 2 — Plugin Marketplace

Inside Claude Code:

/plugin marketplace add justn-hyeok/harness-for-yall

Then install what you need:

/plugin install dev-pipeline@justn-harness
/plugin install review-pipeline@justn-harness
/plugin install fe-experts@justn-harness
/plugin install be-experts@justn-harness
/plugin install explore-team@justn-harness

Option 3 — Shell script

git clone https://github.com/justn-hyeok/harness-for-yall.git
cd harness-for-yall
chmod +x install.sh && ./install.sh

Architecture

                    ┌─────────────┐
                    │  User Task  │
                    └──────┬──────┘
                           │
              ┌────────────┼────────────┐
              ▼            ▼            ▼
        ┌──────────┐ ┌──────────┐ ┌──────────┐
        │ Specific │ │ Generic  │ │ Analysis │
        │ fe / be  │ │   dev    │ │ explore  │
        └────┬─────┘ └────┬─────┘ └────┬─────┘
             │             │             │
             ▼             ▼             ▼
        ┌──────────────────────────────────────┐
        │            review-pipeline           │
        │  screener×3 → moderator → judge      │
        └──────────────────────────────────────┘

Routing Rules

1. Specificity first — fe-* / be-* over dev-* when the task is clearly frontend or backend
2. Chainable — explore → dev → review for complex workflows
3. Skip when overkill — One-line fixes and simple questions don't need a harness

Model Strategy

Agent	Model	Why
explore-scout	opus	Orchestration quality matters for architecture analysis
Everything else	sonnet	Cost efficiency — sonnet handles implementation well

Structure

.claude-plugin/
  marketplace.json        # Plugin Marketplace catalog
plugins/
  dev-pipeline/           # 5 agents, 1 skill
    .claude-plugin/plugin.json
    agents/               # dev-planner, dev-frontend, dev-backend, dev-reviewer, dev-qa
    skills/dev-feature/
  review-pipeline/        # 5 agents, 1 skill
    agents/               # review-screener-{1,2,3}, review-moderator, review-judge
    skills/review-code/
  fe-experts/             # 5 agents, 5 skills
    agents/               # fe-architect, fe-implementer, fe-styler, fe-perf, fe-tester
    skills/               # fe-component, fe-page, fe-refactor, fe-review, fe-test
  be-experts/             # 7 agents, 5 skills
    agents/               # be-architect, be-implementer, be-validator, be-resilience, be-provider, be-security, be-tester
    skills/               # be-api, be-mcp-server, be-pipeline, be-llm-integration, be-observability
  explore-team/           # 4 agents, 3 skills
    agents/               # explore-scout, explore-hypothesizer, explore-evidence, explore-synthesizer
    skills/               # explore-investigate, explore-quick, explore-hypothesis
bin/install.mjs           # npx CLI
package.json              # npm config
install.sh                # Shell installer
uninstall.sh              # Shell uninstaller

Customization

Each agent is a standalone .md file with YAML frontmatter. Edit freely:

---
name: fe-architect
description: "React/Next.js component architecture"
model: sonnet          # change to opus if you want
tools:
  - Read
  - Glob
  - Grep
---

Skills are in SKILL.md format inside named folders. Add your own by creating plugins/<team>/skills/<name>/SKILL.md.

License

MIT