D

claude-api

by来自 anthropics/skills · view source ↗查看源 ↗

Autonomous Coding自主编程see source见源仓库18,325 tok/call18,325 token/次

Skill briefSkill 简述

|- Reference for the Claude API / Anthropic SDK — model ids, pricing, params, streaming, tool use, MCP, agents, caching, token counting, model migration. TRIGGER — read BEFORE opening the target file; don't skip because it "looks like a one…提供 Claude API / Anthropic SDK 的参考文档，涵盖模型 ID、价格、参数、流式传输、工具调用、MCP、agent、缓存、token 计数与模型迁移，规定在打开目标文件前必读，不可因其看似简单而跳过。

Review findings · 8评测结果 · 8

• Description doesn’t say WHEN to use itdescription 没说清什么时候该用它medium中危A good description names the trigger (e.g. "when reviewing a PR"). Without it the skill mis-fires or never fires.好的 description 会点明触发场景（比如"审查 PR 时"）。少了它，skill 要么乱触发、要么根本不触发。
• Trigger is too broad触发条件太宽泛medium中危Words like "whenever", "any", "all" make it fire everywhere and fight other skills for the same task — the #1 cause of conflicts.像 "whenever", "any", "all" 这样的词会让它到处触发，和别的 skill 抢同一个任务——这是冲突的头号原因。
• Bloated: high token cost内容冗长 · token 成本偏高medium中危~18,325 tokens on every single call. Compress instructions and trim examples.每次调用都要多花约 18,325 个 token。精简指令、删减示例就能省下来。
• References API keys / env config引用 API key / 环境变量配置medium中危Reads an API key or env var — normal for API clients. Ensure keys come from the runtime (not hardcoded); a leak risk only if sent to an untrusted endpoint.读取 API key 或环境变量——对 API 客户端来说很正常。确保 key 来自运行时（别硬编码在文件里）；只有把它发往不可信的地址时才有泄露风险。
• Contains curl / wget network command包含 curl / wget 网络命令medium中危Accesses the network from the shell. If the target is untrusted it can pull malware or exfiltrate data.从 shell 直接访问网络。如果目标不可信，可能被用来拉取恶意程序或外传数据。
• Reads credentials and makes network calls — verify the destination既读凭证又发网络请求——请核实目标地址medium中危The normal shape of an API client (your key → the API). Static analysis sees the capability, not the intent — confirm the endpoint is one you trust; it is a data-leak risk only if it is not.API 客户端本来就长这样（你的 key → 对应的 API）。静态分析只看得到"能做什么"，看不到"想干什么"——确认这个地址是你信得过的；只有它不可信时才谈得上泄露风险。
• Description too longdescription 太长low低危The description is read on every trigger check; overly long ones waste tokens and dilute the signal.每次判断要不要触发都会读一遍 description；写太长既浪费 token，又冲淡了关键信号。
• Repeated content内容重复low低危~2 duplicate lines detected; de-duping is free token savings.检测到约 2 行重复内容；去重就是白捡的 token。

Conflicts with · 1与之冲突 · 1

If you have both installed, they compete for the same triggers — the top cause of an agent picking the wrong skill.如果两个都装了，它们会抢同一批触发词 —— 这是 agent 选错 skill 的头号原因。