anthropic-claude-code-agent-skills
Current Claude Code instructions, 22 engineering skills, a TypeScript mod, and a plain concise output style.
- 评测生成时间(北京时间)
- 本报告引擎
- v3.16.0
- 当前引擎
- v3.17.3
本报告与当前引擎使用不同规则;原分数不会自动更新,不同版本的分数不宜直接对比。
进入后确认来源与额度,提交才会创建任务。
代码版本未记录或记录无效,无法核对本报告的 commit 与送评文本;不会用当前项目代码回填历史记录。
综合采用结论
存在需要人工复核的风险或证据不足
- 基础评测完成+25/25确定性评分与静态安全扫描已完成
- README 有效证据+13/255,333 个去重后的有效字符
- 独立证据来源+20/205 类非重复证据,重复文件不叠加
- 仓库元数据+10/10已取得仓库状态与采用数据
- 活跃记录+5/5已取得最近提交时间
- AI 复核+15/15已完成结构化 AI 证据复核
本报告评了什么
综合分用于静态采用筛选,不是任务成功率,也不代表使用 Skill 后的能力提升。
- 静态检查范围
- 记录 19 个文件 · 165,615 字符
- 只检查采集到的受限文本与已知风险模式;未读取的脚本、依赖和运行环境不在结论内。
- AI 内容复核
- 已使用
- 复核送评文本的说明和设计,不是安装测试或实际任务执行。记录缺失时不推断已使用或未使用。
- 真实任务效果
- 本报告未提供实测证据
- 没有使用/不使用 Skill 的任务对照、重复运行或耗时与 token 成本结果。采用前仍需在隔离环境验证自己的任务。
脚本引用覆盖
脚本引用覆盖记录不可得;不推断没有引用或已经检查。
可观测性技能工作流
SKILL.md 的 Workflow 为单一任务的连续 8 步处理,且 Checklist 与参考文件构成前后依赖,适合流程图
左右滑动查看完整图示
- • SKILL.md Workflow 第 1 步:Find the logging library, its configuration, the OpenTelemetry setup or agent
- • SKILL.md Workflow 第 6 步:Run the code with a console exporter or local stdout and read the real output
- • SKILL.md Workflow 第 8 步:Review against the checklist
- 通过: Agent Skills 格式校验 18/18 个通过
- 通过: 18 个有效 Skill 含可核实的指令步骤或示例
- 质量评审选取 claude/skills/adding-observability/SKILL.md;其余 17 个仅做格式扫描
- SKILL.md 的 Workflow 列出 8 步,从查找日志库到按 Checklist 复核
- Checklist 要求“No secret, token, session ID, connection string, key, card data, or personal data appears in logs, spans, baggage, metrics, or exception messages”
- Names and attributes 节规定自定义属性用反向域名前缀如 com.acme.order.tier,禁止以 otel. 开头
- README WARNING 段说明安装器添加 "permissions": { "defaultMode": "bypassPermissions" } 与 skipDangerousModePermissionPrompt
- README 说明 mods 通过 CLAUDE_CODE_PLUGIN_DIRS 加载,需 Claude Code v2.1.280 或更高
- 问题与用途描述
- 有效 README
- 安装或接入步骤
- 可执行示例
- 通过: Agent Skills 格式校验 18/18 个通过
- 发现提示词或系统信息提取意图
- claude/skills/planning-implementation/SKILL.md:description 未清楚说明何时使用该 Skill
- README 警告安装器会写入 permissions.defaultMode=bypassPermissions 与 skipDangerousModePermissionPrompt
- brief-guard 在 bypassPermissions 下放行除 AskUserQuestion/EnterPlanMode/ExitPlanMode 外的所有权限请求
- 技能本身未给出可直接运行的代码示例,示例集中在 examples.md 且未在输入中展示
为已有服务补充 OpenTelemetry 日志、指标、追踪与健康端点、审查现有遥测中的敏感数据泄露与日志注入风险、统一团队属性命名、基数控制与错误记录规范
无法接受默认 bypassPermissions 权限模式的生产环境、需要开箱即用可运行代码示例而非规范清单的场景
也有自己的公开项目?先看完证据,再用当前规则生成独立报告。
评测我的项目 →prompt-extractionclaude/skills/managing-dependencies-and-ci/SKILL.md:19high confidence- **Install scripts and weight.** Read any install or postinstall scripts and look for downloads, obfuscation, or exfiltration. Notice how many indirect dependencies it pulls into 修复:移除提示词提取逻辑,并增加敏感上下文不可输出的边界说明。
- 01移除提示词提取逻辑,并增加敏感上下文不可输出的边界说明。
- 02修复 claude/skills/planning-implementation/SKILL.md:description 未清楚说明何时使用该 Skill
方法、证据与局限展开收起
GitHub Repository API
19 个文件 · 165,615 字符
v3.16.0 · AI 内容复核:已使用(deepseek-flash)
- 静态评测不会安装或执行项目代码
- 安全扫描基于高信号文件与已知模式,不能替代人工审计
- 流行度只反映采用程度,不代表安全或工程质量
- 发现 18 个 Skill;质量复核只评审 claude/skills/adding-observability/SKILL.md,其余 17 个仅做格式扫描
30 天热度趋势
README
Claude Code instructions and skills
A portable configuration for Claude Opus 5.5, Sonnet 5.5, Haiku 5.5, and Fable 5.1, synchronized with the current local setup.
The payload contains 41 engineering skills, the Concise Personal output style, the personal-instructions mod written in TypeScript, and portable settings. No custom subagent definitions or global CLAUDE.md file are installed.
The style asks for plain, natural answers in the user's language. It preserves concise complete sentences, useful lists and tables, and the user's punctuation requirements. Short answers do not reduce the completeness of research or implementation.
Claude Code loads the output style selected in settings through its standard mechanism. The mod preserves that mechanism, loads operating rules from Markdown files, and removes competing tone, communication, narration, and proactive delegation instructions. It puts selected built-in tools and Claude MCP tools behind ToolSearch without changing their input schemas or blocking calls. T3 and third-party MCP loading preferences are preserved. MCP server instructions are separate from deferred tool definitions and are not deferred by this setting.
Intermediate messages are limited to requested status, material blockers or changes, findings that change the next step, and approaches the user needs to understand beforehand. Messages outside thinking blocks follow the language of the user's request. Internal reasoning and thinking blocks have no language requirement. The mod removes the separate fixed-language reminder and selected conflicting Claude Docs instructions. It does not filter generated answers or add recurring reminders.
Install
Requires a Claude Code build supporting mods and TypeScript hooks. The current mod was tested with Claude Code 2.1.314. Windows installers require PowerShell 5.1 or PowerShell 7. The macOS and Linux installer also requires Python 3 for settings processing.
powershell -NoProfile -ExecutionPolicy Bypass -File .\scripts\install.ps1
sh scripts/install.sh
Replace mode backs up and replaces custom skills, agents, output styles, and mods. It removes a legacy global CLAUDE.md file so old instructions do not compete. Credentials, sessions, projects, plugins, synced skills, and unrelated files remain in place. Existing settings are preserved, missing defaults are added, and the active style and mod location are updated. Both installation modes remove the retired skillOverrides setting after backing it up.
The portable settings enable bypassPermissions and skip its confirmation prompts. Review settings.json before installation. The default mod path uses ~/.claude/mods/personal-instructions and the installer resolves custom destinations on the host.
Start a new Claude Code session after installation. See installation details for merge mode and custom destinations.
Update
Run scripts/update.ps1 in PowerShell or sh scripts/update.sh on macOS or Linux. A Git checkout can instead use git pull --ff-only followed by the installer. Updates reuse the last installation mode.
Contents
Account state, sessions, project data, caches, retired files, and cloud-synced skills are excluded. Telegram notifications remain available for future repository updates. This is an independent configuration, not an official Anthropic distribution. Prompt instructions guide behavior and do not guarantee compliance on every response.