Guide · Agent skill
How to use the Harness agent-team designer skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
Harness is an agent skill for Claude that designs a team of specialized AI agents for a specific project. It writes each agent's role into a file, writes the skills those agents follow, and records how to call the setup later. This guide explains what the skill does, how to install it, how to trigger it, and what to check before you accept its output.
Harness agent-team designerrevfactory/harness · Apache-2.0 Download the skill Quick start
- 1Download the ZIP. Download the Harness skill ZIP from this page.
- 2Unzip it into a skills folder. Unzip into ~/.claude/skills/ for personal use, or into your project's .claude/skills/ folder for one project. Alternatively, upload the ZIP under Settings → Capabilities → Skills in the Claude apps.
- 3Confirm your environment. Make sure Claude Code has the multi-agent features the skill references, and that your project folder is writable.
- 4Describe your project. Ask in plain language, for example '이 프로젝트에 맞는 에이전트 팀 만들어줘', and state the domain and goal.
- 5Review the proposed plan. The skill checks existing files and proposes an execution mode and team layout. Confirm or adjust it before files are created.
- 6Review the generated files. Read the agent definitions, skills and CLAUDE.md entries, and check tool permissions and model choices.
What is an agent skill, and what does Harness make the agent do?
An agent skill is a folder with a SKILL.md file that tells Claude how to handle a certain kind of task. When a request matches the skill's description, Claude loads the instructions and follows them.
Harness is a skill for planning a multi-agent setup. Its stated job is to design a harness suited to a project, define the specialized agents, and create the skills those agents will use. Each agent definition says who does the work, and each skill says how the work is done.
- Defines each agent's role in a file under .claude/agents/
- Chooses how the agents run together: a scripted workflow, persistent collaborating agents, or one-off sub-agents
- Picks a model per task based on complexity, duration, autonomy and response speed
- Writes skills under .claude/skills/{name}/SKILL.md
- Checks an existing setup for mismatches between agents, skills and CLAUDE.md
- Records invocation conditions and a change history in CLAUDE.md so the setup can be reused in new sessions
Install the Harness skill
Download the ZIP file from this page. You can install it in one of three ways, depending on where you want it to be available.
Option 1 is your personal skills folder, which makes the skill available across projects. Option 2 is a single project's .claude/skills/ folder, which keeps it with that project. Option 3 is uploading the ZIP in the Claude apps under Settings → Capabilities → Skills.
- Personal: unzip the file into ~/.claude/skills/
- Project: unzip the file into your project's .claude/skills/ folder
- Claude apps: go to Settings → Capabilities → Skills and upload the ZIP
Check the requirements before you start
The skill is written for Claude Code with its multi-agent features. The skill itself refers to the Agent tool, the Workflow tool, SendMessage and TaskCreate, so these need to be available in your environment. If they are not, the skill's workflow-based and persistent-agent modes will not run as described.
You also need a project folder where the skill can write .claude/agents/, .claude/skills/ and CLAUDE.md. According to the skill's metadata, it does not run scripts and does not use the network.
- Claude Code with multi-agent features (Agent, Workflow, SendMessage, TaskCreate)
- A project folder where .claude/agents, .claude/skills and CLAUDE.md can be written
- Scripts: none run by this skill, per its metadata
- Network: not used, per its metadata
How to trigger Harness with natural requests
You do not need a special command. Describe what you want in plain language that matches the skill's trigger phrases. The skill is meant to activate for requests to build or design a harness, or to create an agent team.
For maintenance, use requests about checking or syncing an existing setup. For reviewing a run and feeding lessons back into the setup, the skill points to a separate skill called harness:evolve.
- '하네스 구성해줘' (set up a harness)
- '하네스 구축해줘' (build a harness)
- '하네스 설계' (harness design)
- '하네스 엔지니어링' (harness engineering)
- '에이전트 팀 만들어줘' (make an agent team)
- '하네스 점검' or '하네스 감사' (check or audit the harness)
- '하네스 현황' (harness status)
- '에이전트/스킬 동기화' (sync agents and skills)
- For retrospectives on a finished run, use harness:evolve
Example: designing a team for a codebase
Suppose you want an automated review setup for a project. You could ask: '이 프로젝트에 맞는 에이전트 팀 만들어줘' (make an agent team that fits this project). The skill first checks whether .claude/agents/, .claude/skills/ and CLAUDE.md already exist, then reads the codebase to understand the stack and main modules.
Next it proposes an execution mode and a team layout, and asks you to confirm before creating files. Expect the output to be agent definitions, skills and a CLAUDE.md entry, not just a chat answer.
- Existing agents and skills are checked first to avoid duplicates
- Execution mode is chosen per workflow: scripted, persistent, or sub-agent
- Models are assigned per task and the reason is noted in comments
Review the output and stay in control
Generated agents and workflows can run commands and spend significant resources. Read each agent file, skill and orchestrator before you rely on them. Do not accept generated files unreviewed if you do not understand agent configuration.
The skill says the Workflow tool requires your explicit consent. Its default is a small number of agents, and it only scales up when you ask for a broad investigation. Keep that limit in mind when you request thorough work.
- Read every generated file in .claude/agents/ and .claude/skills/
- Check tools lists: read-only reviewers should not have Edit or Write
- Ask for a smaller team if the default feels too large
- Use the skill's audit requests to catch mismatches between files and CLAUDE.md
Tips for getting useful results
Give the skill a concrete project goal and a clear list of the work types involved, such as creating, verifying, editing or analysing. The skill uses these to decide which agents are needed.
Use the skill for structured setups, not for simple one-off jobs. The setup process adds overhead that a single prompt does not need. Once a setup exists, keep CLAUDE.md current so later sessions can call it correctly.
- State the project's domain, tech stack and goal in your first message
- Name the work types you need: generation, verification, editing, analysis
- Ask for an audit when files and CLAUDE.md may have drifted apart
- Review and test generated agents before running them on real work
What people use it for
- Setting up a multi-agent workflow for a codebase, with separate agents for analysis, generation and verification
- Extending an existing agent and skill configuration with a new specialist without creating duplicates
- Auditing a project's .claude/agents, .claude/skills and CLAUDE.md for mismatches
- Syncing agent and skill definitions after files have been changed by hand
Frequently asked questions
How do I install the Harness agent skill?
Download the ZIP from this page and unzip it into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder. You can also upload the ZIP in the Claude apps under Settings → Capabilities → Skills.
What do I need to run the Harness skill?
You need Claude Code with the multi-agent features the skill references, namely the Agent tool, Workflow tool, SendMessage and TaskCreate. You also need a project folder where .claude/agents, .claude/skills and CLAUDE.md can be written.
Is the Harness skill free to use?
This page does not state pricing. The source lists the licence as Apache-2.0, so check the licence text in the repository for the terms that apply to you.
Does Harness run scripts or access the network?
According to the skill's metadata, it runs no scripts and uses no network. The agents and workflows it generates can still run commands and use significant resources, so review them before use.
Why won't the skill trigger for my request?
Use wording close to the trigger phrases, such as '하네스 구성해줘', '하네스 설계' or '에이전트 팀 만들어줘'. Also confirm the skill is installed in the location Claude Code reads, and that the required multi-agent features are available.
How does Harness compare with writing agents by hand?
Harness plans the team first, including roles, execution mode, model choice and skills, then writes the files and checks for overlap with existing ones. Writing by hand gives you full control but no structured plan. For a single small task, a plain prompt is usually simpler.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Harness suits developers and teams who want a structured, auditable multi-agent setup for a codebase, or who need to extend or check an existing agent and skill configuration. It is not worth the overhead for simple one-off tasks, and it is a poor fit if you would accept generated agent files without reading them, since those files can run commands and use significant resources.
Download the skill