III. How to find an agent

1. Skills and harnesses

Skills

A skill is a set of operating instructions organized in an agreed format and loaded for a task. It can include reference materials and programs. Its value is that software can discover, load and reuse a working method. Agent Skills overview

How is it different from AGENTS.md?

The content can be identical. The difference is where it is placed and when it loads:

• AGENTS.md provides instructions for a project or directory, such as its structure, editing rules and test commands. Supporting software loads it according to directory scope. AGENTS.md
• A skill packages a type of task separately, with a name and description. The software first tells the model which skills exist, then loads the detailed instructions when a task needs them. How loading works

Under the Agent Skills specification, create a folder containing SKILL.md. Start the file with name and description, followed by the instructions. The name must match the folder. Here is a minimal example:

return-support/
└── SKILL.md

Contents of SKILL.md:

---
name: return-support
description: Check a shop's return conditions and draft customer replies. Use for inquiries about return requests.
---

# Handle a returns inquiry

1. Get the current return policy, the time the customer received the order and details of product use.
2. Ask for information that affects eligibility; do not guess.
3. Apply the policy to explain whether a request can be submitted and what review follows.
4. Distinguish eligibility to apply from a guaranteed refund. Do not add promises absent from the policy.

This folder is already a skill resource in the required format. Add references/, scripts/ or other directories when you need supporting rules, examples or programs, and specify which files to read or run. An AGENTS.md file does not automatically become a skill. A SKILL.md without helper programs can still be a skill. Format specification

A skill file does not run itself. Supporting software must discover it, expose its name and description to the model, and place the full instructions in context when selected. The model can select it from the task, or the user can request it explicitly. Integration guide

In Claude Code, for example, put this folder under the project's .claude/skills/, giving .claude/skills/return-support/SKILL.md. Start a message with /return-support, followed by the customer's question and materials, to load it explicitly. Claude Code skill directories and invocation

Appearing in the skill list confirms discovery. An explicit invocation or read record helps confirm loading. Check the actual steps and results to establish whether it was followed correctly.

Text instructions cannot guarantee perfect compliance. Enforce mandatory restrictions in programs and tools: grant no write permission for read-only work; make the registration interface reject requests without an order number when one is required. The skill explains how to work; the runtime checks which operations may happen.

The basic deliverable is the whole folder. Include referenced materials and programs, preserve relative paths, and share it by copying the directory or publishing a repository.

For upload to Claude on the web, create a ZIP with this structure:

return-support.zip
└── return-support/
    ├── SKILL.md
    └── references/   Include supporting materials if the skill uses them.

In Windows PowerShell, run this from the skill folder's parent directory:

Compress-Archive -LiteralPath .\return-support -DestinationPath .\return-support.zip

In Claude, go to Customize → Skills, add a skill, upload the ZIP and enable it. Claude Code can use the folder directly without a ZIP. Packaging requirements, uploading and enabling

Harnesses

A harness is the program that runs an agent, organizing model calls, tool execution, context and execution controls. It runs a loop between the model and its tools. When the model requests an order lookup, the harness passes the request to the query tool and returns the result to the model. Anthropic describes the core as a loop that calls the model and passes on tool requests. Anthropic: runtime structure

A harness and a workflow can be the same thing. “Harness” emphasizes the program running the model; “workflow” emphasizes task steps, branches and order.

The core is the execution loop between model and tools. A complete harness may combine these capabilities:

File access, skill loading, task decomposition and subagents can be added to this foundation. Deep Agents provides several of these capabilities. Other systems may supply model services, file systems and databases; the harness organizes them. Deep Agents: components and capabilities

2. How to find agents

Chinese-language resources

3. How to find skills

4. How to find harnesses

5. How to find workflows