Agent team harness that teaches Claude Code your team's rules — not the other way around.
Claude Code is powerful, but out of the box it doesn't know your coding standards, architecture patterns, review checklists, or deployment workflows. Claude Harness Kit bridges that gap with a specialist agent team, structured workflows, and convention-first design.
Your conventions. Your agents. Your workflow. Claude just follows.
Same code-reviewer agent behaves differently per project:
- In a Spring Boot project → checks jOOQ usage, hexagonal boundaries, REST Docs
- In a React project → checks hooks rules, component structure, CSS modules
How? Agents read your CLAUDE.md and docs/standards/ — not hardcoded rules.
Most AI coding tools assume you're starting from scratch. Real teams aren't.
Claude Harness Kit is designed for projects that already have:
- Established coding standards and architecture patterns
- Existing CI/CD pipelines and build systems
- Team conventions that must be followed, not reinvented
- Issue trackers (Jira, GitHub Issues) with ongoing work
/task analyze → Convert specs to structured markdown (PPT, Google Docs, Confluence)
/task plan → Generate planning doc (plan-analyst agent)
/task design → Generate design doc (design-architect agent)
/task impl → Implement code
/task review → Parallel review (3 agents: code / architecture / test)
/task done → Build verification + completion summary
/task e2e → End-to-end API tests
/jira-task start → Create branch, transition Jira issue
/jira-task plan~done → Same phases + Jira status sync & comments
/jira-task report → Sprint status report
GitHub Issues and Linear integrations are planned.
Works inside or outside the workflow:
| Skill | Description |
|---|---|
/analyze-spec |
Convert planning docs to structured MD (PPT, Google Slides/Docs/Sheets, Confluence) |
/db |
Local infrastructure management (Docker, DB, cache) |
/server |
Service start/stop/build/status |
/e2e-test |
E2E API testing (curl + Playwright) |
/propagate-convention |
Propagate new ADR/rules to all agent/skill files |
/cleanup-worktree |
Git worktree cleanup |
Customizable templates in skills/examples/:
| Example | Description | Tech Stack |
|---|---|---|
domain-model |
DDD domain model design guide | Spring Boot + jOOQ |
dto-design |
DTO layer separation guide | Spring Boot |
port-adapter |
Hexagonal Architecture guide | Spring Boot + jOOQ |
query-audit |
DB query quality audit | jOOQ |
type-split |
Wide table split + sealed interface | Java 17+ |
test-convention |
Test convention guide | JUnit 5 |
domain-refactor |
Domain code refactoring | DDD + Clean Arch |
Each example has [CUSTOMIZE] markers for easy adaptation to your tech stack.
| Hook | Event | Description |
|---|---|---|
session-logger.sh |
SessionStart, UserPromptSubmit, PostToolUse | Audit trail — logs session start, prompts, git commands to .worktree.log |
guard-merge.sh |
PreToolUse(Bash) | Merge guard — blocks merge if current branch != baseBranch |
notify.sh |
Notification | Desktop notification on task completion (macOS/Linux/WSL) |
Agent = Role (what to do) ← shared (agents/)
+ Standards (how to judge) ← your project (see below)
7 specialist agents work as a team:
| Agent | Role | Model |
|---|---|---|
pm |
PM/orchestrator — Jira management, team coordination | opus |
plan-analyst |
Requirements analysis, task breakdown, risk identification | opus |
design-architect |
Technical design, sequence diagrams, no code | opus |
developer |
Implementation — follows existing patterns 100% | sonnet |
code-reviewer |
Coding standards, security, complexity review | opus |
architecture-reviewer |
Layer structure, module boundaries, pattern consistency | opus |
test-reviewer |
Test conventions, coverage, naming | sonnet |
convention-keeper |
Propagates new ADR/rules to all agent/skill files | sonnet |
Agents read your project standards in this order:
1. CLAUDE.md ← Project overview, core patterns (~50 lines)
2. .claude/task-conventions.md ← Per-phase conventions (~50 lines)
3. docs/standards/*.md ← Detailed standards (if any)
4. Existing code (Glob/Grep) ← Reference patterns
5. (nothing found) → General principles + ask the user
CLAUDE.md stays under 200 lines by splitting into 3 layers:
| File | Role | Lines |
|---|---|---|
CLAUDE.md |
Project overview + pointers | ~50 |
.claude/task-conventions.md |
Per-phase conventions (plan/design/impl/test/review) | ~50 |
docs/standards/*.md |
Detailed standards (coding, API, error handling, etc.) | as needed |
New project? Just write ~30 lines in CLAUDE.md. Conventions can be added later.
# One-liner — run from your project root
bash <(curl -sL https://raw.githubusercontent.com/AnSungWook/ClaudeHarnessKit/main/setup.sh)Or clone manually:
git clone https://github.com/AnSungWook/ClaudeHarnessKit.git /tmp/claude-harness-kit
cd /path/to/your/project
bash /tmp/claude-harness-kit/setup.shThe installer will:
- Copy shared skills (analyze-spec, db, server, e2e-test, propagate-convention)
- Choose a workflow (task or jira-task) + copy 8 specialist agents
- Copy hooks (session-logger, guard-merge, notify)
- Generate
config.yaml,settings.json,.mcp.jsonwith your project settings - Create convention directories (
docs/adr/,docs/standards/with templates) - Create artifact directories (
docs/specs,plan,design,review,test,reports) - Generate
CLAUDE.md+task-conventions.mdtemplates
Edit .claude/skills/config.yaml for your project:
project:
name: "My Project"
root: "/path/to/project"
workflow:
type: task # task | jira-task
server:
modules:
- name: api
port: 8080
start: "./gradlew bootRun"
health: "http://localhost:8080/actuator/health"
infra:
type: docker # docker | none
commands:
up: "docker-compose up -d"See config.yaml for full options.
claude-harness-kit/
├── config.yaml # Project settings (server, infra, paths)
├── setup.sh # Installer script
│
├── agents/ # Specialist agent team (8 agents)
│ ├── pm.md # PM / orchestrator (opus)
│ ├── plan-analyst.md # Planning analysis (opus)
│ ├── design-architect.md # Design (opus)
│ ├── developer.md # Implementation (sonnet)
│ ├── code-reviewer.md # Coding standards review (opus)
│ ├── architecture-reviewer.md # Architecture review (opus)
│ ├── test-reviewer.md # Test conventions (sonnet)
│ └── convention-keeper.md # Rule propagation (sonnet)
│
├── skills/
│ ├── task/ # Standalone task workflow (no issue tracker needed)
│ ├── analyze-spec/ # Spec analysis
│ ├── server/ # Server management
│ ├── db/ # Infrastructure management
│ ├── e2e-test/ # E2E testing
│ ├── propagate-convention/ # ADR/rule propagation to agent/skill files
│ ├── examples/ # Tech-specific example skills (7 examples)
│ └── workflows/
│ └── jira-task/ # Jira integration (13 sub-skills)
│
├── commands/
│ └── cleanup-worktree.md # Worktree cleanup
├── hooks/
│ ├── guard-merge.sh # Merge guard hook
│ ├── session-logger.sh # Session audit trail (3 events)
│ └── notify.sh # Desktop notification
│
├── templates/ # Templates for CLAUDE.md, conventions, settings, MCP
│ ├── CLAUDE.md.template
│ ├── task-conventions.md.template
│ ├── settings.json.template # Hooks + deny list + MCP env vars
│ ├── .mcp.json.template # Jira, PostgreSQL, Playwright MCP config
│ └── docs/
│ ├── adr/ # ADR template + guide
│ ├── standards/ # Coding, API, testing standard templates
│ └── artifacts/ # Workflow artifact directory guide
└── docs/ # Design docs and usage guides
├── GUIDE-task-workflow.md
├── DESIGN-task-workflow.md
└── token-optimization-guide.md # Token usage optimization guide
/task analyze → plan → design → impl → review → done → e2e
Each phase uses specialized agents (Opus for planning/review, Sonnet for implementation). Review runs 3 agents in parallel: code, architecture, and test.
/jira-task start → plan → design → impl → test → review → done
Same workflow as task, plus automatic Jira status transitions, issue comments, branch/PR management, and sprint reporting. Requires one of:
| MCP Server | Install |
|---|---|
| sooperset/mcp-atlassian (community, 4.8k stars) | claude mcp add atlassian -- uvx mcp-atlassian --jira-url URL ... |
| atlassian/atlassian-mcp-server (official) | npx @anthropic-ai/create-mcp |
13 sub-skills — install the plugin:
claude plugin install jira-integration@jira-claude-code-integrationSee Task Workflow Guide for details.
| Requirement | Used by | Install |
|---|---|---|
| python-pptx | analyze-spec | pip install python-pptx |
| LibreOffice | analyze-spec | brew install --cask libreoffice |
| Docker | db | Docker Desktop |
| Playwright MCP | e2e-test | Add @playwright/mcp to .mcp.json |
| Jira MCP | jira-task | claude mcp add atlassian ... |
| Google MCP | analyze-spec | Add to .mcp.json |
Manages your CLAUDE.md lifecycle — captures session learnings and audits quality.
claude plugin install claude-md-management@claude-plugins-official| Command/Skill | Purpose | When |
|---|---|---|
/revise-claude-md |
Capture session learnings into CLAUDE.md | End of each session |
claude-md-improver |
Audit CLAUDE.md quality (100-point rubric) | Periodic maintenance |
Why it matters for the harness:
- Session learnings accumulate in CLAUDE.md → next session doesn't need to rediscover patterns
- Quality audit keeps CLAUDE.md concise → reduces per-turn token overhead
- Distinguishes
CLAUDE.md(team-shared) vs.claude.local.md(personal)
- Task Workflow Guide — Usage, convention paths, new project setup
- Task Workflow Design — Architecture, detailed specs
- Token Optimization Guide — Token usage optimization for harness workflows
MIT