Best Skill Recommendations
Primary role: Evaluate and recommend skills already discovered by an upstream skill search. Only when no upstream results exist does this skill independently search clawhub, then apply the same evaluation logic.
Prerequisites
Before this skill can operate, confirm the following:
- -
clawhub CLI is installed and available in the agent's PATH. This skill issues clawhub search, clawhub list, clawhub install, and clawhub uninstall commands. If the binary is absent, all discovery and install steps will fail. - Authentication: Run
clawhub login in advance. The clawhub CLI stores credentials in its own default config path (managed by the CLI, not by this skill). No additional environment variables are required by this skill. - Permissions in scope: This skill will only:
1. Read the installed skill list via
clawhub list (read-only).
2. Run
clawhub install <skill> or
clawhub uninstall <skill> —
only after explicit user confirmation at the Pre-Install Gate (Step 5).
It will not access other system files, credentials, or APIs beyond the clawhub CLI.
Mandatory Store Policy
- 1. Prefer upstream results first. If any upstream skill search has already returned candidates, use those directly — do not re-search.
- Only when no upstream candidates exist: search
clawhub directly. - Never claim exclusivity; both public and private registries are valid.
- Before install, always summarize source, version, and notable risk signals.
- If a fresh search is needed, run
clawhub search <keywords> and report output before evaluating.
Auto-Trigger and Collaboration
- - Preferred entry: triggered after an upstream skill search has already produced a candidate list. Reuse those results directly.
- Standalone entry: triggered when the user asks to recommend/install/compare skills but no upstream results exist. In this case, independently search
clawhub and produce candidates before evaluating. - Never re-search if usable upstream candidates are already available.
Workflow
0) Check for Upstream Candidates (first)
Before doing anything else:
- - If an upstream skill search has already returned a candidate list → skip to Step 2 using those results.
- If no upstream candidates exist → proceed to Step 1.
1) Self-Discover (only when no upstream results)
Clarify the user's need:
- - target task(s)
- priority (speed/stability/features/safety)
- constraints (region, cost, runtime)
Then search via clawhub:
clawhub search <keywords>
Return the command output and build a candidate list annotated with source and version.
2) Evaluate Candidates
Present the candidate list (from upstream or self-discovered), annotated with source and version.
3) Evaluate Installed Skills and Compatibility
Enumerate installed skills by running:
clawhub list
Parse the output to get the current installed skill set. Compare each installed skill against candidates:
- - overlap: full / partial / complementary
- conflict risk: command/workflow collision, duplicated automation, behavior mismatch
- coexist feasibility: high / medium / low
4) Recommend Replace vs Coexist
Per candidate, output one decision:
- - Replace existing skill(s)
- Coexist with boundaries
- Do not install
Include reasons and trade-offs.
5) Pre-Install Gate (required)
Before any install action, present:
- - source (clawhub)
- version
- notable risk signals — check for any of the following:
- install count < 10 (low adoption)
- last update > 6 months ago (low maintenance signal)
- requires broad file system or network permissions
- published < 2 weeks ago (very new, unvetted)
- no author URL or verified identity
Then explicitly ask for user confirmation.
6) Install Execution
- - INLINECODE15
- If replace approved: uninstall old first, then install new.
- If coexist: install and provide usage boundary guidance.
7) Post-Install Report
Return:
- - install/uninstall result
- final active skill set
- why this is best for the user goal
- follow-up checks
Output Format
Candidate Summary
- - Name
- Purpose
- Source
- Version
- Install Count (if available)
- Link
Decision Summary
- - Best choice (Top 1)
- Alternatives (Top 2/3)
- Replace/Coexist decision
- Why not other options
Pre-Install Confirmation
"Planned action:
. Source: clawhub. Version: <...>. Risks: <...>. Proceed?"
Guardrails
- - Never install without explicit confirmation.
- Prefer fewer, higher-confidence recommendations over long noisy lists.
- If search quality is poor, refine keywords and re-run search before recommending.
最佳技能推荐
主要职责: 评估并推荐上游技能搜索已发现的技能。仅当没有上游结果时,此技能才会独立搜索 clawhub,然后应用相同的评估逻辑。
前提条件
在此技能运行之前,请确认以下事项:
- - clawhub CLI 已安装 并在代理的 PATH 中可用。此技能会执行 clawhub search、clawhub list、clawhub install 和 clawhub uninstall 命令。如果二进制文件缺失,所有发现和安装步骤都将失败。
- 身份验证: 提前运行 clawhub login。clawhub CLI 将其凭据存储在自己的默认配置路径中(由 CLI 管理,而非此技能)。此技能不需要额外的环境变量。
- 权限范围: 此技能仅会:
1. 通过 clawhub list 读取已安装的技能列表(只读)。
2. 运行 clawhub install
或 clawhub uninstall — 仅在预安装门控(第 5 步)获得用户明确确认后。
它不会访问除 clawhub CLI 之外的其他系统文件、凭据或 API。
强制商店策略
- 1. 优先使用上游结果。 如果任何上游技能搜索已返回候选,直接使用这些结果 — 不要重新搜索。
- 仅当没有上游候选时:直接搜索 clawhub。
- 绝不声称排他性;公共和私有注册表均有效。
- 安装前,始终总结来源、版本和显著风险信号。
- 如果需要全新搜索,运行 clawhub search 并在评估前报告输出。
自动触发与协作
- - 首选入口: 在上游技能搜索已生成候选列表后触发。直接重用这些结果。
- 独立入口: 当用户要求推荐/安装/比较技能但不存在上游结果时触发。在这种情况下,在评估前独立搜索 clawhub 并生成候选。
- 如果已有可用的上游候选,绝不重新搜索。
工作流程
0) 检查上游候选(优先)
在执行任何其他操作之前:
- - 如果上游技能搜索已返回候选列表 → 跳至第 2 步,使用这些结果。
- 如果没有上游候选 → 继续执行第 1 步。
1) 自我发现(仅当没有上游结果时)
明确用户需求:
- - 目标任务
- 优先级(速度/稳定性/功能/安全性)
- 约束条件(区域、成本、运行时)
然后通过 clawhub 搜索:
clawhub search <关键词>
返回命令输出并构建一个标注了来源和版本的候选列表。
2) 评估候选
展示候选列表(来自上游或自我发现),标注来源和版本。
3) 评估已安装技能及兼容性
通过运行以下命令枚举已安装技能:
clawhub list
解析输出以获取当前已安装的技能集合。将每个已安装技能与候选进行比较:
- - 重叠:完全 / 部分 / 互补
- 冲突风险:命令/工作流冲突、重复自动化、行为不匹配
- 共存可行性:高 / 中 / 低
4) 推荐替换或共存
针对每个候选,输出一个决策:
包含理由和权衡。
5) 预安装门控(必需)
在任何安装操作之前,展示:
- - 来源(clawhub)
- 版本
- 显著风险信号 — 检查任何以下情况:
- 安装次数 < 10(采用率低)
- 上次更新距今 > 6 个月(维护信号低)
- 需要广泛的文件系统或网络权限
- 发布时间距今 < 2 周(非常新,未经审查)
- 没有作者 URL 或已验证身份
然后明确请求用户确认。
6) 安装执行
- - clawhub install
- 如果批准替换:先卸载旧技能,再安装新技能。
- 如果共存:安装并提供使用边界指导。
7) 安装后报告
返回:
- - 安装/卸载结果
- 最终活跃技能集合
- 为何这对用户目标最佳
- 后续检查
输出格式
候选摘要
决策摘要
- - 最佳选择(Top 1)
- 备选方案(Top 2/3)
- 替换/共存决策
- 为何不选其他选项
预安装确认
计划操作:<安装/替换/共存>。来源:clawhub。版本:<...>。风险:<...>。是否继续?
护栏
- - 未经明确确认绝不安装。
- 优先选择数量更少、置信度更高的推荐,而非冗长嘈杂的列表。
- 如果搜索质量差,优化关键词并重新运行搜索后再进行推荐。