← Back to Skills Marketplace
zhelunsun

Harness

by Julian Zhelun Sun · GitHub ↗ · v2.4.0 · MIT-0
cross-platform ✓ Security Clean
25
Downloads
0
Stars
0
Active Installs
1
Versions
Install in OpenClaw
/install skill-factory-claw
Description
Agent-first engineering knowledge base distilled from OpenAI's Harness Engineering post. Use this skill when the user asks about: agent-first development, ag...
README (SKILL.md)

\r \r

Harness: Agent-First Engineering Knowledge Base\r

\r

核心定位\r

\r 本 Skill 是对 OpenAI 团队在 Harness Engineering 实践中积累的第一手智能体工程经验的系统化蒸馏。\r \r 背景:5个月内,3名工程师仅用提示(无人工编码)、借助 Codex 智能体,构建了约 100 万行代码 的产品。\r

  • ~1,500 个 Pull Requests\r
  • 平均每人每天 3.5 个 PR 的吞吐量\r
  • 完成时间约为传统人工编写的 1/10\r \r ---\r \r

使用场景\r

\r 当用户问到以下问题时,加载本 Skill 的 references/ 文档作为上下文来回答:\r \r

  • 如何设计面向智能体的代码仓库结构?\r
  • AGENTS.md 应该怎么写?\r
  • 智能体工程中的上下文管理策略?\r
  • 如何让应用对智能体"可读"(可观测)?\r
  • 如何用架构约束替代微观管理?\r
  • 如何处理智能体引入的技术债务?\r
  • 人类工程师在智能体优先团队中的角色是什么?\r
  • 智能体的自主水平如何随系统成熟度提升?\r \r ---\r \r

Knowledge References\r

\r

harness/\r
├── SKILL.md                            # 本文件(入口)\r
└── references/\r
    ├── harness-engineering.md          # OpenAI 原文完整蒸馏(主文档)\r
    ├── core-principles.md              # 10大核心原则速查表\r
    └── extended-sources.md             # 扩展信息源索引(Anthropic/GitHub/大V等)\r
```\r
\r
加载知识时:\r
1. 优先读取 `references/core-principles.md` 快速定位\r
2. 需要深度背景时读取 `references/harness-engineering.md`\r
3. 需要扩展视角时读取 `references/extended-sources.md`\r
\r
**多源整合**:本 skill 不仅包含 OpenAI Harness Engineering,还整合了 Anthropic Claude Code、GitHub Copilot Workspace、Aider 等业界最佳实践,持续追踪智能体工程前沿。\r
\r
---\r
\r
## Workflow\r
\r
1. **识别用户问题的核心维度**:上下文管理 / 架构约束 / 可观测性 / 技术债务 / 人类角色 / 合并策略 / 文档体系\r
2. **读取对应 references 文档**\r
3. **结合用户具体场景**,从 Harness 经验中提炼可操作的建议\r
4. **输出时注明来源**:明确这是 OpenAI Harness Engineering 实践中验证过的方法\r
\r
---\r
\r
## 快速摘要(无需读文件即可回答的核心原则)\r
\r
| 领域 | 黄金法则 |\r
|------|---------|\r
| 情境管理 | AGENTS.md ≤ 100行,作为目录指向深层文档;不要巨型配置文件 |\r
| 架构 | 严格分层(Types→Config→Repo→Service→Runtime→UI),通过 linter 自动强制 |\r
| 可读性 | UI + 日志 + 指标对智能体直接可读;接入 DevTools 协议 |\r
| 技术债务 | 黄金原则 + 循环清理 = 持续垃圾回收;小额偿还,不累积 |\r
| 合并策略 | 纠错成本低 > 等待成本高;减少阻塞门 |\r
| 文档 | 仓库是唯一记录系统;知识必须 push 进 repo 才对智能体可见 |\r
| 人类角色 | 设计环境 + 明确意图 + 构建反馈回路;不写代码,写约束 |\r
| 技术选型 | 偏好 API 稳定、可组合、在训练集中表现良好的技术栈 |\r
| 架构约束时机 | 严格分层约束是早期先决条件,不要等到百人团队再考虑 |\r
| 自我审查闭环 | 智能体在本地审核自身变更(Ralph Wiggum 循环),直到所有审阅人满意才合并 |\r
Usage Guidance
Install only if you intend to use ClawHub/OpenClaw registry tooling and are comfortable granting it the credentials needed for publishing, syncing, GitHub workflows, or staff moderation. Review any command that deletes, restores, publishes, comments on PRs, or changes user roles before approving it.
Capability Assessment
Purpose & Capability
The repository and local skill files consistently describe OpenClaw skill/package registry work, CLI install/sync/publish flows, Convex development guidance, UI proofing, and staff moderation; the higher-impact actions fit that stated purpose.
Instruction Scope
Several repo-local skills can guide code changes, PR comments, UI proof publication, migrations, or moderation actions, but the instructions generally require explicit targets, confirmation, verification, and normal authenticated API paths.
Install Mechanism
The package metadata shows normal Bun/npm workspace usage, declared CLI binaries, build/test scripts, and no hidden postinstall persistence; the root preinstall only enforces Bun usage.
Credentials
Network access, API tokens, GitHub/Convex/OpenAI configuration, local skill scanning, and optional telemetry are expected for a registry/CLI/development tool and are documented rather than hidden.
Persistence & Privilege
The artifacts include authenticated registry mutations such as publishing, soft delete/restore, package operations, and staff moderation, but they are scoped to user/admin roles and include prompts, reasons, audit/logging expectations, or reversibility where appropriate.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install skill-factory-claw
  3. After installation, invoke the skill by name or use /skill-factory-claw
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v2.4.0
- Added new skill "harness," a distilled agent-first engineering knowledge base based on OpenAI's Harness Engineering post. - Covers best practices and principles for building AI agent workflows, codebase structuring, context management, error tolerance, observability, and technical debt handling. - Provides quick-access guidelines, detailed references, and integration of multiple industry sources (OpenAI, Anthropic, GitHub Copilot Workspace, etc.). - Outlines user workflow: identify user need, consult targeted references, and deliver actionable advice citing proven Harness Engineering methods. - Includes concise, table-form core rules for rapid user reference. - Offers bilingual support (English & Chinese terminology) for relevant agent engineering topics.
Metadata
Slug skill-factory-claw
Version 2.4.0
License MIT-0
All-time Installs 0
Active Installs 0
Total Versions 1
Frequently Asked Questions

What is Harness?

Agent-first engineering knowledge base distilled from OpenAI's Harness Engineering post. Use this skill when the user asks about: agent-first development, ag... It is an AI Agent Skill for Claude Code / OpenClaw, with 25 downloads so far.

How do I install Harness?

Run "/install skill-factory-claw" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is Harness free?

Yes, Harness is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does Harness support?

Harness is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created Harness?

It is built and maintained by Julian Zhelun Sun (@zhelunsun); the current version is v2.4.0.

💬 Comments