SmallClaw 与 OpenClaw 的 Skill 使用对比 SmallClaw vs OpenClaw Skill Usage

一句话结论 One-line conclusion

SmallClaw 和 OpenClaw 都使用 Markdown skill 作为工作流层，但 SmallClaw 更偏向“先审后用、运行时治理更强”，OpenClaw 更偏向“CLI / registry / install 的技能管理方式”。 SmallClaw and OpenClaw both use Markdown skills as a workflow layer, but SmallClaw leans toward “review first, then use” with stronger runtime governance, while OpenClaw leans toward a CLI / registry / install model for skill management.

SmallClaw 如何使用 Skill How SmallClaw uses skills

SmallClaw 把 skill 当成本地 Markdown 工作流说明，而不是可直接执行的代码模块。它的核心是文件、registry 和 runtime 三层协作。 SmallClaw treats skills as local Markdown workflow guidance rather than directly executable code modules. Its core is the collaboration of file, registry, and runtime layers.

1. Skill 存在哪里 1. Where skills live

SmallClaw 自动从以下目录加载 skill： SmallClaw automatically loads skills from the following directories:

• <workspace>/.smallclaw/skills
• ~/.smallclaw/skills
• ~/.openclaw/skills

最后一个目录是 OpenClaw 兼容路径。SmallClaw 会把这里的 skill 读取进自己的 registry 和 runtime 索引里。 The last directory is the OpenClaw compatibility path. SmallClaw reads skills from there into its registry and runtime index.

2. 如何创建 2. How to create them

在 SmallClaw 里，最直接的创建方式是打开 Skills 页面，然后使用内置模板新建 skill。通常会填写 name、description、triggers、enabled 和 instructions，保存后就会写成 Markdown 文件。 In SmallClaw, the most direct creation flow is to open the Skills page and create a skill with the built-in template. You usually fill in name, description, triggers, enabled, and instructions, then save it as Markdown.

3. 如何编辑 3. How to edit them

已加载的 skill 可以在 Skills 页面里打开并编辑。保存后，内容会回写到对应的 Markdown 文件。这意味着 SmallClaw 的 skill 是“文件为主”的：文件是主要内容载体，数据库注册负责记录来源、hash、信任状态、启用状态和使用状态，runtime 负责把匹配到的 skill 注入到本次模型请求中。 Loaded skills can be opened and edited in the Skills page. Saving writes back to the corresponding Markdown file. This means SmallClaw is file-first: the file is the main content carrier, the database registry records source, hash, trust state, enabled state, and usage state, and runtime injects matched skills into the current model request.

4. 如何导入外部 skill 4. How to import external skills

SmallClaw 允许导入 OpenClaw 风格的 skill，也支持从 GitHub 预览和安装公开的 skill 仓库。导入流程通常是：打开 Skills，选择 Browse OpenClaw Skills 或 Import Skill From GitHub，先预览 metadata 和文件内容，再选择安装目标并确认。 SmallClaw supports importing OpenClaw-style skills and also previewing and installing public skill repositories from GitHub. The usual flow is: open Skills, choose Browse OpenClaw Skills or Import Skill From GitHub, preview metadata and file contents first, choose an install target, and confirm.

需要注意的是，Browse OpenClaw Skills 需要开启 Permissions > Network Access；公开 skill 会先预览再安装确认；外来 skill 默认更保守，通常会进入 Pending Review。 Keep in mind that Browse OpenClaw Skills requires Permissions > Network Access; public skills are previewed before install confirmation; and external skills start conservatively, usually in Pending Review.

5. 如何触发 5. How to trigger them

SmallClaw 支持两种触发方式：显式触发可以写成 @skill(Your Skill Name) 或 #skill:your-skill-name；自动匹配则依赖 prompt 里出现 skill 名称或 trigger 关键词。如果你希望可预测，显式触发是更稳的方式。 SmallClaw supports two trigger styles: explicit triggers such as @skill(Your Skill Name) or #skill:your-skill-name, and automatic matching when the skill name or trigger keywords appear in the prompt. If you want predictable behavior, explicit triggering is the safer choice.

6. 如何进入运行时 6. How they enter runtime

SmallClaw 不是把 skill 当成独立插件代码去运行，而是把它注入当前任务的 runtime prompt。也就是说，skill 先被加载，registry 决定它是否可信、是否启用，AgentRuntime 在发给模型前做匹配，命中的 skill 作为 Local Skill Context 注入。 SmallClaw does not run skills as standalone plugin code. It injects them into the current runtime prompt. In practice, the skill is loaded first, the registry decides whether it is trusted and enabled, AgentRuntime matches it before the request is sent, and the matched skill is injected as Local Skill Context.

7. 如何确认生效 7. How to verify it worked

发送任务后，可以留意 Agent Console 下方的 Skills: ...、状态栏里的 matched skills，以及 Skills 页里的 Approved / Pending Review / Blocked / Enabled 状态。如果没有命中，通常先检查 skill 是否已经加载、是否启用、外来 skill 是否已批准，以及 prompt 是否用了显式 @skill(...)。 After sending a task, look for Skills: ... under Agent Console, matched skills in the status line, and Approved / Pending Review / Blocked / Enabled in the Skills page. If nothing matches, check whether the skill is loaded, whether it is enabled, whether the external skill is approved, and whether you used the explicit @skill(...) form.

8. 外来 skill 的安全策略 8. External skill safety

SmallClaw 对外来 skill 采用默认保守策略：来自 ~/.openclaw/skills 的内容会被识别为外来来源，openClawCompatible 格式默认更偏复审，只有 approved 且 enabled 的 skill 才能进入注入链路。一旦已批准的外来 skill 文件内容变化，就会因为 hash 变化回到 Pending Review。这让 SmallClaw 更像“先审后用”。 SmallClaw uses a conservative default for external skills: content under ~/.openclaw/skills is classified as external, openClawCompatible files are review-oriented by default, only approved and enabled skills can enter the injection path, and if an approved external skill changes on disk, it falls back to Pending Review. That makes SmallClaw more “review first, then use.”

OpenClaw 如何使用 Skill How OpenClaw uses skills

OpenClaw 的公开文档把 skill 组织成更偏 CLI / registry / install 的使用方式。它更像是在管理可安装资源，而不是先进入一个图形化创作器。 OpenClaw’s public documentation frames skills more as a CLI / registry / install workflow. It feels more like managing installable resources than starting from a graphical creator.

1. 如何管理 1. How they are managed

OpenClaw 提供类似 openclaw skills search、install、update、list、info、check 这样的命令。这种方式更像“命令行管理安装包和本地 skill 状态”。 OpenClaw provides commands such as openclaw skills search, install, update, list, info, and check. This feels more like command-line management of installable skill assets and local skill state.

2. 如何获取 2. How they are obtained

OpenClaw 的官方文档通常围绕本地 skills、ClawHub 上的安装和更新、skill 目录和兼容配置、skill eligibility / check 来组织用户流程。 OpenClaw documentation usually focuses on local skills, ClawHub installation and updates, skill directories and compatibility settings, and skill eligibility / check workflows.

3. 如何使用 3. How they are used

OpenClaw 的使用方式通常更接近：先搜索 skill、再安装 skill、然后更新 skill、查看本地 skill 列表、检查 skill 状态。从用户心智上看，它更偏“把 skill 当作可安装可更新的资源”。 OpenClaw’s workflow is usually closer to searching for a skill, installing it, updating it, listing local skills, and checking skill status. From the user’s mental model, skills are manageable resources that you install and inspect.

两者的核心差异 Core differences

实际怎么选 Which one to use

如果你在 SmallClaw 里工作，自己常用的流程放进 Workspace，经常复用的个人流程放进 User，需要兼容 OpenClaw 文件结构时放进 OpenClaw，导入陌生 skill 先预览再确认，想稳定命中则优先用 @skill(...)。 If you are working in SmallClaw, put recurring personal workflows into Workspace, reusable personal workflows into User, use OpenClaw when you need compatibility with OpenClaw file structure, preview unfamiliar skills before confirming them, and use @skill(...) when you want stable matching.

如果你在 OpenClaw 的语境里工作，先用搜索和安装把 skill 加进本地环境，再用 list、info、check 确认状态，把 skill 当成可管理资源，而不是一次性复制进去的文本片段。 If you are working in the OpenClaw context, use search and install to bring skills into the local environment, then use list, info, and check to confirm status. Treat skills as manageable resources rather than copied text snippets.

迁移建议 Migration advice

如果你是从 OpenClaw 切到 SmallClaw，先把 OpenClaw 风格的 skill 放到 ~/.openclaw/skills，打开 SmallClaw 的 Skills 页面，检查它是否被识别为 OpenClaw-compatible，确认是否需要复审，用 @skill(...) 触发一次来确认命中。 If you are moving from OpenClaw to SmallClaw, put the OpenClaw-style skill into ~/.openclaw/skills, open SmallClaw Skills, check whether it is recognized as OpenClaw-compatible, confirm whether it needs review, and trigger it once with @skill(...) to confirm the match.

如果你是从 SmallClaw 迁回 OpenClaw，保持 Markdown 文件结构清晰，保留 name、description、triggers 和 instructions，尽量把 skill 写成能被目录和 CLI 识别的形式，再用 OpenClaw 的 list / check 流程确认兼容性。 If you are moving from SmallClaw back to OpenClaw, keep the Markdown structure clean, preserve name, description, triggers, and instructions, write the skill so it can be recognized by directory and CLI workflows, and use OpenClaw list / check to confirm compatibility.

小结 Summary

SmallClaw 和 OpenClaw 都把 skill 当作工作流层，但 SmallClaw 更强调图形化创作器、目录监听、运行时 prompt 注入、外来 skill 复审，以及 hash 变化后的重新确认；OpenClaw 更强调 CLI 管理、搜索 / 安装 / 更新、以及本地 skill 检查和可用性控制。 SmallClaw and OpenClaw both treat skills as a workflow layer, but SmallClaw emphasizes a graphical skill creator, watched folders, runtime prompt injection, review for external skills, and reapproval after hash changes. OpenClaw emphasizes CLI management, search / install / update, and local skill checks plus availability control.

如果你的重点是“本地 macOS agent 的安全使用”，SmallClaw 的方式更保守；如果你的重点是“通过 CLI 管理一组可安装技能”，OpenClaw 的方式更像一个技能管理工具链。 If your priority is safe local use on a macOS agent, SmallClaw is more conservative. If your priority is managing a set of installable skills through CLI, OpenClaw feels more like a skill management toolchain.