Guide · Agent skill
How to use the Agent Harness Construction agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
Agent Harness Construction is an agent skill, a set of instructions that Claude can load when a task calls for it. This one gives design guidance for the tools an AI agent can call, the format of their results, how to recover from errors, and how to keep the context small. It is aimed at developers building or tuning their own agent systems, and at teams whose agents loop, retry too often or fail to finish tasks. After this guide you will know how to install it, how to ask for it in plain language, and what kind of advice to expect.
Agent Harness Constructionaffaan-m/ECC · MIT Download the skill Quick start
- 1Download the ZIP. Download the skill ZIP file from this page.
- 2Unzip the skill. Unzip the file. It contains SKILL.md.
- 3Choose a location. Place the folder in ~/.claude/skills/ for personal use, or in .claude/skills/ inside a project.
- 4Or upload in the Claude apps. In the Claude apps, go to Settings → Capabilities → Skills and upload the ZIP.
- 5Describe your agent problem. Ask in plain language about tool design, error handling, retries or context size.
- 6Apply and measure. Make the suggested changes and track completion rate and retries per task.
What does the skill make the agent do?
The skill is a set of design principles, written in Japanese, that the agent applies when you ask it to plan, call tools, recover from errors or converge on a finished result. It states that output quality depends on four things: the quality of the action space (the tools available), the quality of observations (what tools return), the quality of recovery, and the quality of the context budget.
It does not run anything itself. It does not perform tasks, call services or change your files. It only gives advice that the agent uses while helping you design or review an agent system.
- Tool design: stable, explicit tool names and narrow, input-schema-first definitions.
- Result format: a consistent shape with status, summary, next_actions and artifacts.
- Error recovery: root-cause hints, safe retry instructions and explicit stop conditions.
- Context budget: a minimal, stable system prompt and references to files instead of long inline text.
- Architecture choice: when to use ReAct, function calling, or the recommended hybrid of ReAct planning with typed tool execution.
Install the skill from a ZIP file
Download the ZIP file from this page and unzip it. The package contains a single file, SKILL.md. Place the unzipped skill folder in one of two locations: your personal skills folder at ~/.claude/skills/, or a project-specific folder at .claude/skills/ inside your project.
If you use the Claude apps instead of Claude Code, you can upload the skill under Settings → Capabilities → Skills. Use the same ZIP file.
- Personal install: ~/.claude/skills/ (available across your projects).
- Project install: .claude/skills/ inside the project folder (shared with that project).
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP.
How to trigger the skill with natural requests
You do not need a special command. Describe the problem in plain language and mention the agent or tools you are working on. The skill is meant for planning, tool calls, error recovery and convergence to completion, so requests on those topics are the best match.
Good requests name the situation: what the agent does, where it goes wrong, and what you are building.
- Design a tool set for my support-ticket agent. Keep the tool names clear and the outputs consistent.
- My agent keeps retrying the same failed API call. Review its error handling and suggest stop conditions.
- Review our agent's tool responses and suggest a format with status, summary and next actions.
- Our agent's system prompt is getting long. Suggest how to keep the context budget small.
Examples of what the guidance covers
The skill's guidance on granularity is specific. It recommends micro-tools for high-risk operations such as deployments, migrations and permission changes. It recommends medium-sized tools for common edit, read and search loops. It recommends macro-tools only when round-trip overhead is the dominant cost.
Each tool response should include status (success, warning or error), a one-line summary, next_actions for possible follow-ups, and artifacts such as file paths or IDs. Every error path should include a root-cause hint, a safe retry instruction and an explicit stop condition.
- Granularity: micro-tools for deployments and permissions, medium tools for everyday edits and searches.
- Anti-patterns to check: overlapping tools with similar meaning, opaque outputs with no recovery hint, error-only outputs with no next step, and irrelevant references that overload context.
What you need and what it does not do
The skill lists no requirements. It does not run scripts and does not use the network. You do not need to install any software beyond what is needed to use the agent skill itself.
The file also lists metrics to track, including completion rate, retries per task, pass@1 and pass@3, and cost per successful task. It does not include benchmarks or evidence that the advice improves these numbers, so measure your own system before and after applying changes.
- Requirements: none listed.
- Scripts: none.
- Network access: none.
- Metrics: listed as things to track, with no benchmark results included.
Tips for getting useful advice
Bring concrete material. Paste your tool list, a sample response, or an error log, and ask the agent to review it against the principles. Advice on real examples is more useful than general statements.
Ask for a before-and-after comparison, then test the change against your own task set. Because the skill is in Japanese, you may get better results by asking in Japanese, or by asking for the answer in the language your team uses.
- Paste real tool definitions and responses rather than describing them.
- Ask for a checklist you can apply to each tool.
- Track completion rate and retries before and after changes.
What people use it for
- Designing a set of tools for an internal support agent, with clear names and no overlapping tools.
- Standardising the response format of tools so the agent always sees status, summary and next actions.
- Diagnosing an agent that loops or retries the same failing call, and adding explicit stop conditions.
- Reducing context size by moving long guidance into files the agent reads only when needed.
- Choosing between ReAct, function calling or a hybrid architecture for a new agent project.
Frequently asked questions
How do I install the agent-harness-construction skill?
Download the ZIP from this page and unzip it. Put the folder in ~/.claude/skills/ for personal use or in .claude/skills/ inside a project. In the Claude apps, you can instead upload the ZIP under Settings → Capabilities → Skills.
Does this skill need any requirements or software?
The skill lists no requirements. It does not run scripts and does not use the network, so it adds no extra software dependencies.
Is the agent-harness-construction skill free?
The source is licensed under MIT. The page does not list any price for the skill, so check the repository for the licence terms that apply.
Will this skill automate my tasks?
No. It is a set of design principles and does not perform tasks itself. It gives the agent advice on designing tools, result formats and error recovery.
Does it prove that the advice improves agent performance?
No. It lists metrics such as completion rate and retries per task, but it includes no benchmarks or evidence. Measure your own agent before and after making changes.
What should I use instead if I need a ready-made automation?
This skill is not a tool for end users who want a finished automation. If you need an agent that performs work, look for a tool or program built for that task, since this skill only gives design advice.
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
Choose this skill if you are designing tools, response formats or error handling for your own AI agent, or diagnosing agents that loop or fail to finish. Skip it if you want an automation that does the work for you, or if you need measured proof of results, because it offers design principles without benchmarks. Its guidance is written in Japanese, so check that it fits your team.
Download the skill